Application of Artificial Intelligence in Food Industry—a Guideline

Artificial intelligence (AI) has embodied the recent technology in the food industry over the past few decades due to the rising of food demands in line with the increasing of the world population. The capability of the said intelligent systems in various tasks such as food quality determination, control tools, classification of food, and prediction purposes has intensified their demand in the food industry. Therefore, this paper reviews those diverse applications in comparing their advantages, limitations, and formulations as a guideline for selecting the most appropriate methods in enhancing future AI- and food industry–related developments. Furthermore, the integration of this system with other devices such as electronic nose, electronic tongue, computer vision system, and near infrared spectroscopy (NIR) is also emphasized, all of which will benefit both the industry players and consumers.


Introduction
Artificial intelligence (AI) is defined as a field in computer science that imitates human thinking processes, learning ability, and storage of knowledge [1,2]. AI can be categorized into two types which are strong AI and weak AI. The weak AI principle is to construct the machine to act as an intelligent unit where it mimics the human judgments, while the strong AI principle states that the machine can actually represent the human mind [3]. However, strong AI does not exist yet and the study on this AI is still in progress. The gaming industry, weather forecasting, heavy industry, process industry, food industry, medical industry, data mining, stem cells, and knowledge representation are among the areas that have been utilizing AI methods [4][5][6][7][8][9][10][11]. AI has a variety of algorithms to choose from such as reinforcement learning, expert system, fuzzy logic (FL), swarm intelligence, Turing test, cognitive science, artificial neural network (ANN), and logic programming [3]. The alluring performance of AI has made it the most favorable tool to apply in industries including decision making and process estimation aiming at overall cost reduction, quality enhancement, and profitability improvement [7,12].
As the population in the world is rising, food demand is predicted to rise from 59 to 98% by 2050 [13]. Thus, to cater for this food demand, AI has been applied such as in management of the supply chain, food sorting, production development, food quality improvement, and proper industrial hygiene [14][15][16]. Sharma stated that the food processing and handling industries are expected to grow about CAGR of 5% at least until 2021 [15]. ANN has been used as a tool in aiding real complex problem solving in the food industry according to Funes and coworkers [17], while based on Correa et al., the classification and prediction of parameters are simpler when using ANN, which leads to higher usage demand of ANN over the past years [18]. Besides, FL and ANN have also acted as controllers in ensuring food safety, quality control, yield increment, and production cost reduction [19,20]. AI technologies have also known to be beneficial in food drying technology and as process control for the drying process [21][22][23].
Previous studies have shown many usages of AI in food industries focusing on individual target and aims. A study has been conducted on the various ANN applications in food process modeling where it has only highlighted the food process modeling using ANN [24]. Apart from that, the implementation of AI such as ANN, FL, and expert system in food industries have been reviewed but specifically focusing on the drying of fresh fruits [23]. A review has been conducted on how food safety has been one of the main concerns in the food industry which leads to the development of smart packaging systems to fulfill the requirements of the food supply chain. Intelligent packaging monitors the condition of foods to give details on the quality of the food during storage and transportation [25]. Another study reviewed on intelligent packaging as a tool to minimize food waste where about 45 recent advances in the field of optical systems for freshness monitoring have been reported. Meat, fish products, fruits, and vegetables were covered in the study as they are the most representative fields of application [25]. Few different studies have been conducted on intelligent packaging, and these studies proved that the usage of intelligent packaging systems plays an important role in the food factory in the context of the food chain as they are able to monitor the freshness of food products and crops [23,[26][27][28][29][30].
There are also several other studies that have been conducted on the application of AI and sensors in food; however, the coverage is rather limited. Therefore, a comprehensive review that assembles all AI applications in the food industry as well as its combinations with appropriate sensor will be a great advantage, all of which are unavailable as to the knowledge of the author. Such review will assist in gathering the advantages, limitations, and methodologies as a one-stop guideline and reference for food industry players, practitioners, and academicians. To be exact, different types of AI and their recent application in food industries will be highlighted which comprises several AI techniques including expert system, fuzzy logic, ANN, and machine learning. In addition, the integration of AI with electronic nose (E-nose), electronic tongue (E-tongue), near infrared spectroscopy (NIRS), and computer vision system (CVS) is also provided. This paper is organized as follows. The introduction of AI is explained in the first section followed by the application of different types of AI in the food industry. Following that, the fusion of the AI with the external sensors in the food industry is presented. In the latter part, a critical review is conducted where discussion on the main application of the AI algorithms in the food industry is carried out. A flowchart is presented to assist the researchers on establishing the most appropriate AI model based on their specific case study. Then, the trends on the application of AI in the food industry are illustrated after that section. Finally, a brief conclusion is discussed in this paper.

AI in Food Industry
The application of AI in the food industry has been growing for years due to various reasons such as food sorting, classification and prediction of the parameters, quality control, and food safety. Expert system, fuzzy logic, ANN, adaptive neuro-fuzzy inference system (ANFIS), and machine learning are among the popular techniques that have been utilized in the food industries. Prior to AI implementation, studies related to food have been going on over the years to educate the public about food knowledge as well as to improve the final outcomes related to food properties and the production of foods [31][32][33][34][35][36]. A lot of benefits can be obtained by using the AI method, and its implementation in the food industry has been going on since decades ago and has been increasing till today [37-39, 31, 32]. Nevertheless, this paper will focus on the application of AI in food industries from the year 2015 onwards since tremendous increase and innovation are seen in the implementation recently. It is worth noting that several methods such as partial least square, gastrointestinal unified theoretical framework, in silico models, empirical models, sparse regression, successive projections algorithms, and competitive adaptive reweighted sampling which have been used for prediction and enhancement of the food industries are not discussed here; instead it is narrowed down to the wide application of AI in the food industry.

Knowledge-based Expert System in Food Industry
The knowledge-based system is a computer program that utilizes knowledge from different sources, information, and data to solve complicated problems. It can be classified into three categories which are expert systems, knowledge-based artificial intelligence, and knowledge-based engineering. The breakdown of the knowledge-based system is presented in Fig. 1. The knowledge-based expert system which is widely used in the industries is a decisive and collective computer system that is able to imitate the decision-making ability of human expert [40]. It is a type of knowledge-based system that is known as among the first successful AI models. This system depends on experts for solving the complicated issues in a particular domain. It has two sub-systems, which are knowledge base and inference engine. The facts about the world are stored in the knowledge base, and the inference engine represents the rules and conditions regarding the world which are usually expressed in terms of the IF-THEN rules [41]. Normally, it is able to resolve complicated issues by the aid of a human expert. This system is based on the knowledge from the experts. The main components of the expert system (ES) are human expert, knowledge engineer, knowledge base, inference engine, user interface, and the user. The flow of the expert system is shown in Fig. 2.
The food industry has been utilizing ES for various objectives as this system is proven to be useful especially in the decision-making process. The knowledge-based expert system has been applied in white winemaking during the fermentation process for the supervision, intelligent control, and data recovery [42]. Apart from that, a web-based application was developed by implementing the ES to calculate the nutritional value of the food for the users, and the development of ES was able to help the SMIs in obtaining the details required for the qualification in obtaining the food production certificates [43]. Food safety is very important in the food industry,thus, the application of ES that is linked closely to food safety has been used extensively ranging from process design, safety management, quality of food, and risk assessment [44]. Furthermore, a prototype information technology tool and guidelines with corrective actions that considered ES in the model were developed for the food industry where few essential factors such as food safety, nutrition, quality, and cost were studied [45]. In addition, a digital learning tool, namely, MESTRAL, was developed to assist people in food processing by using models developed from research in food science and technology and simulators. This tool is based on the knowledge engineering and reflected real applications which can be mapped with the system scale and knowledge frameworks [46]. A comprehensive review was conducted by Leo Kumar on the application of the knowledge-based expert system in manufacturing planning. The paper has also discussed the utilization of ES in decision making in three wide areas which are the process planning activities, diverse applications, and manufacturing planning [41]. Moreover, Table 1 gathers some of the recent application of ES in the food industry ranging from the raw material to the final production as well as the food safety.

Fuzzy Logic Technique in the Food Industry
Fuzzy logic (FL) was first introduced by Zadeh in 1965 based on the impeccable capability of human intellect in decision making and unraveling the imprecise, uncertain, and ambiguous data while solving problems [47,48]. The fuzzy set theory is recognized in such a manner that an element belongs to a fuzzy set with a certain degree of membership which has a real number in the interval [0, 1] [49]. FL models consist of several steps which are fuzzification, inference system, and defuzzification process [50,51]. Fuzzification is a process where the crisp value is converted into a degree of membership and yields the fuzzy input sets. The corresponding degree in the membership functions is normally between 0 and 1. [52]. There are a variety of membership functions to choose from, whereby the commonly used ones are triangular, Z-shaped, S-shaped, trapezoidal, and Gaussian-shaped [52]. The inference system is where the fuzzy input is being translated to get output by using the fuzzy rules. The fuzzy rules are known as IF-THEN rules where it is written such IF premise, THEN consequent whereby the IF comprises input parameters and THEN is the output parameters [53]. The inference system consists of the style which is either the Mamdani or Takagi-Sugeno Kang (TSK). Defuzzification is the ultimate phase in the fuzzy logic model where the crisp values are obtained [54]. There are different methods of defuzzification which are center of gravity, mean of maximum, smallest of maximum, largest of maximum, center of maximum, and centroid of area [55].
FL has been long utilized in the industry due to its simplicity and ability to solve problems in a fast and accurate manner. FL has been employed in the food industry in food modeling, control, and classification and in addressing foodrelated problems by managing human reasoning in linguistic terms [56]. The food manufacturing system has improved Food products To monitor and forecast the product quality in the production process Production/food quality (i) An intelligent ES was developed which is able to monitor the product quality indicators and make changes to the existing methods; recommendations for the products and the defects for the final product can be identified by this system Blagoveshchenskiy et al. [138] Fresh food To optimize the distribution networks of the fresh foods

References
Livestock production (milk, meat) To analyze the outcome of various variables on the performance of the livestock production Production/raw material (i) ES can be utilized as a decision support system for livestock producers for identifying the best practices for the livestock which maximizes the production of the meat and milk (ii) The greatest impact on the production is the type of grazing being fed to the cattle as the diet affects the health of the cow who produces the milk and the meat Vásquez et al. [140] Red wine and rum To forecast the key aroma compounds for foods without using the human olfactory system Sensory evaluation/quality control (i) The developed rapid method system, Sensomics Based Expert System, resulted in a good agreement in the key odorants for the food aroma distillate Nicolotti et al. [141] Rice crops To aid the farmers in making decisions for the rice crops Agriculture/ production (i) The farmers are assisted by the ES in selecting the seeds and tackling the pests and diseases for the rice crops which eventually will improve the production of the rice crops Kharisma et al. [142] Soybean To identify the diseases on soybeans Agriculture (i) The study successfully designed an ES to identify soybean diseases by comparing the accuracy using the frame-based representation and rule-based representation method (ii) Frame-based representation ES has shown a higher accuracy compared to the rule-based ES Rajendra et al. [143] White winemaking To develop a knowledge-based ES for the alcoholic fermentation process of the white winemaking Processing/sustainability (i) A cost-efficiency advanced control system through the knowledge-based ES was developed for the alcoholic fermentation process which was used for the supervision, control, and data recovery software of the bioreactor (ii) It was proven to be applied in winemaking at the industrial scale and can be adjusted for few areas in the food manufacturing sectors Sipos [42] Wine To measure the environmental impact of viticulture at wine estate Sustainability (i) ES integrated with the geographic information system software was able to measure the environmental impact of viticulture in a comprehensive way (ii) The model is said to be an environmental support system in supporting policy and decisionmaking in the management Lamastra et al. [144] by the implementation of the fuzzy logic where about 7% of electricity losses has been reduced compared to the conventional regulation method [57]. Sensory evaluation of the food is also one of the most common parts where FL plays an important role. Furthermore, a quicker solution to problems can be performed by using a system involving fuzzy rules [58]. Table 2 shows previous applications of FL in the food industry and their attributes. From a previous study, FL has been proven to successfully maintain the quality of the foods, and it acts as a prediction tool and control system for food production processes.
ANN Technique in the Food Industry ANN is another AI element, which is also commonly applied in the food industry. ANN is designed to mimic the human brain and be able to gain knowledge through learning and the inter-neuro connections which are known as synaptic weights [59,60]. Gandhi and coworkers have stated that the configuration of ANN is designed in such a way that it will accommodate certain application such as data classification or pattern recognition [61]. According to Gonzalez-Fernandez, ANN is applicable to a different kind of problems and situations, adaptable, and flexible. In addition,  have also stated that ANN is suitable to model most non-linear systems and is adaptable to new situations even though adjustments are needed. Moreover, the most outstanding features of ANN is its non-linear regression [62]. There are several types of ANN including feedforward neural network, radial basis function neural network, Kohonen self-organizing neural network, recurrent neural network, convolutional neural network, and modular neural network [63]. Multilayer perceptron (MLP), radial basis function networks (RBFNN), and Kohonen self-organizing algorithms are the most effective types of NN when it comes to solving real problems [61]. The most common network that is used for prediction and pattern recognition is the multilayer perceptron [18,64,65]. Besides that, ANN learning could be classified into supervised and unsupervised depending on the learning techniques [17]. In general, the structure of ANN consisted of an input layer, hidden layer, and output layer, either single or many layers [66][67][68]. The architecture comprises activation functions, namely, the feed-forward or feedback [69]. The backpropagation learning algorithm is normally used as it is able to minimize the prediction error by feeding it back as an input until the minimum acceptable error is obtained [18]. An additional input known as bias is added to neurons which allows a portrayal of phenomena having thresholds [70,71]. In ANN, the dataset is normally associated with a learning algorithm which trained the network and could be categorized into three groups specifically supervised, unsupervised, and reinforcement learning [72]. Then, the data will undergo training and testing for analyzing the outputs. The general structure for the ANN is shown in Fig. 3. The output data can be calculated by using the equation shown based on Fig. 4. Previous studies have highlighted the utilization of ANN in numerous tasks within the food industry. This includes the assessment and classification of the samples, complex calculation such as heat and mass transfer, and analysis of the existing data for control purposes as well as for prediction purposes which are listed in Table 3. All applications have shown satisfactory performances based on the R 2 values, showing that ANN can provide results in an accurate and reliable manner.

Machine Learning Techniques
Machine learning (ML) is known to be the subset of AI [73,74]. It is a computer algorithm that advances automatically with experiences. ML can be classified into three broad categories which are supervised learning, unsupervised learning, and reinforcement learning [11,75]. Supervised learning aims to predict the desired target or output by applying the given set of inputs [76]. On the other hand, unsupervised learning does not have any outputs to be predicted and this method is utilized to classify the given data and determine the naturally occurring patterns [77]. Reinforcement learning is when there is an interaction between the program and the environment in reaching certain goals [78]. Among the known models in machine learning are ANN, decision trees (DT), support vector machines (SVM), regression analysis, Bayesian networks, genetic algorithm, kernel machines, and federated learning [76,79]. ML has been commonly used for handling complex tasks and huge amount of data as well as variety of variables where no preformula or existing formula is available for the problem. Other than that, ML models have the additional ability to learn from examples instead of being programmed with rules [80].
Among the ML methods that are used in the food industry include ordinary least square regression (OLS-R), stepwise linear regression (SL-R), principal component regression (PC-R), partial least square regression (PLS-R), support vector regression (SVM-R), boosted logistic regression (BLR), random forest regression (RF-R), and k-nearest neighbors' regression (kNN-R) [81]. Studies showed that the usage of ML has helped in reducing the sensory evaluation cost, in decision making, and in enhancing business strategies so as to cater users' need [82]. Long short-term memory (LSTM) which is an artificial recurrent neural network has been employed in the food industry as pH detection in the cheese fermentation process [83]. On the other hand, GA has been utilized for finding the optimum parameters in food whereas NN has been occupied to predict the final fouling rate in food processing [84]. ML has shown to be advantageous in predicting the food insecurity in the UK [85]. Apart from Canned food To control sterilization temperature by using fuzzy logic and making online corrections in autoclave operation Mamdani Triangular (i) The sterilizing temperature with an accuracy of ± 0.5 ℃ can be maintained by a fuzzy controller (ii) Batch processing can be completed using the proposed system with less time, steam consumption, and risk of over-sterilization Chung et al. [147] Coffee To determine the suitable process on a dry mill according to customer requirements using an expert system based on fuzzy logic Mamdani Triangular (i) The developed system will be useful for the correct decision process between two different types of coffee (ii) Validation was carried out by comparing the process values by the model with the real process data, and the coefficient of determination obtained was 93% Hernández-Vera et al. [148] Coffee beans To introduce a control system for the roasting machine Mamdani Triangular (i) The consistent roasting level of the beans is able to be produced by the proposed model Harsawardana et al. [149] Cupcakes To rank the cupcakes according to their quality attributes Mamdani Triangular (i) The system was able to determine the best condition for their cupcakes with respect to their sensory attributes (ii) The ranking of the quality attributes was able to be performed by the system Singh et al. [150] Dough To implement the FL to act as a controller system in bread making  Pizza production industry To develop a system in order to improve the production system Mamdani Triangular (i) The developed FL control system is able to identify the amounts of workers and ovens needed in pizza production which improves the customer's satisfactory level by reducing the waiting time as well as reducing the wastage Blasi [159] Salt To estimate the production of salt by variables that affect it TSK Triangular (i) By using the Sugeno zero-order model, the time for production of salt could be estimated with a minor error value of 0.0917 Yulianto et al. [160] that, ML has also proven to have predicted the trend of sales in the food industry [86] In addition to that, ML was also able to predict the food waste generated and give an insight to the production system [87]. Major applications of ML in the food industry and its positive highlights are briefly emphasized in Table 4.

Adaptive Neuro Fuzzy Inference System (ANFIS) Techniques
ANFIS is a type of AI where FL and ANN are combined in such a way that it integrates the human-like reasoning style of the FL system with the computational and learning capabilities of ANN [56]. In ANFIS, the learning procedure is transferred from the neural network into the FL system where a set of fuzzy rules with suitable membership functions from the data obtained is developed [88]. Mamat et al. [89] stated that uncertainty data could be processed and gain higher accuracy when ANFIS is applied [89]. Besides, ANFIS is also known as a fast and robust method in solving problems [90]. Not only that, Sharma et al. [91] also claimed that ANFIS has a higher performance compared to other models such as ANN and multiple regression models in their study [91]. ANFIS is a fuzzy reasoning system and combination of the parameters trained by ANN-based algorithms. The fuzzy inference system that is normally used is Takagi Sugeno Kang in the ANFIS model with the feedforward neural network consisting of the learning algorithms [92]. The structure of ANFIS is made up of five layers which are fuzzy layer, product layer, normalized layer, defuzzification layer, and total output layer [93,31,32]. The backpropagation algorithm has been normally applied in the model in order to avoid over-fitting from occurring [92]. A high correlation value (R 2 ) indicates that the developed model has high accuracy and is suitable for industrial applications. The general structure of the ANFIS model is illustrated in Fig. 5. The first layer in ANFIS has nodes that are adjustable, and it is called as the premise parameters [56]. The second layer in ANFIS has fixed nodes, and the output is the product of all incoming signals. Every output node represents the firing strength of the rule. The third layer consists of fixed node labeled as N. The outputs of the third layer are called normalized firing strengths. Every node in the fourth layer is an adaptive node with a node function, and the parameters in this layer are called as the subsequent parameters [56]. The final layer in the ANFIS layer has a fixed single node which calculates the overall output as the summation of all the incoming signals. The calculation involved in each layer is shown below. The output of the ith model in layer 1 is denoted as 0 1 , i.
; w i is the normalized firing strength from layer 3 and.
{p i , q i , r i } is the parameter set of this node. Layer 5: The ANFIS model is attractive enough that it could solve problems related to the food industry, which are complicated, practical, and barely solved by other methods and has been widely used in the food industry for prediction and classification purposes. ANFIS has been applied in various food processing involving recent technology which comprised five main categories which are food property prediction, drying of food, thermal process modeling, microbial growth, and quality control of food as well as food rheology [56]. The utilization of ANFIS in the food industry has been commenced years ago, and Table 5 describes those applications.

Integrating AI with External Sensors for Real-time Detection in Food Industry
FL or ANN is often integrated with several sensors for realtime detection such as electronic nose (E-nose), electronic tongue (E-tongue), machine learning (ML), computer vision system (CVS), and near infrared spectroscopy (NIRS) for real-time detection and to obtain higher accuracy results in a shorter time. These detectors have also combined their elements together for enhancing their accuracy and targeted results. The integration of these sensors with the artificial intelligence methods has been shown quite a number in food industries over the past few years.
Electronic nose also known as E-nose is an instrument created to sense odors or flavors in analogy to the human nose. It consists of an array of electronic chemical sensors where it is able to recognize both simple and complex odors [94]. E-nose has been used in gas sensing where the analysis of each component or mixture of gases/vapors is required. Besides, it plays an important role in the food industry for controlling the quality of the products. Due to its ability to detect complex Total output, Y = F(Y in )   Meat OLS-R, SL-R, PC-R, PLS-R, SVM-R, RF-R, and kNN-R (i) Different kinds of microorganisms causing the beef spoilage could be detected by using the regression ML that obtained the data from five different analytical methods (ii) All the methods were able to predict in all cases accurately with the rank order of RF-R, PLS-R, kNN-R, PC-R, and SVM-R Estelles-Lopez et al. [81] odors, it has been employed as an environment protection tool and detection of explosives materials [95]. An array of nonspecific gas sensors is known to be the main hardware component of E-nose where the sensors will interact with a variety of chemicals with differing strengths. It then stimulates the sensors in the array where characteristic response is extracted known as a fingerprint [94]. The main software component of E-nose is its feature extraction and pattern recognition algorithms where the response is processed, important details are elicited and then chosen. Thus, the software component of the E-nose is greatly important to stimulate its performance. In general, E-nose is divided into three main parts, namely, sample delivery system, a detection system, and a computing system. ANN, FL, and pattern recognitions are the examples of the methodology employed in E-nose [96]. The general system of E-nose is shown in Fig. 6. E-nose has been widely used to aid in both quality control and assurance in the food industries. Wines, grains, cooking oils, eggs, dairy products, meat and dairy products, meat, fish products, fresh-cut and processed vegetables, tea, coffee, and juices have successfully applied e-nose for sampling classification, detection, and quality control. E-nose has successfully classified samples with different molecular compounds [97]. Besides, Sanaeifar et al. have reviewed and confirmed that e-nose was able to detect defects and contamination in foodstuffs [98]. Classification and differentiation of different fruits have also determined by using e-nose [99]. A review has been conducted on the application of the E-nose for monitoring the authenticity  Layer 5 x y x y of food [100]. Adding to this, Mohamed et al. have carried out a comprehensive review on the classification of food freshness by using e-nose integrated with the FL and ANN method [101]. Recent application of e-nose with computing methods involving AI in food industries is shown in Table 6.
Electronic tongue (E-tongue) is an instrument that is able to determine and analyze taste. Several low-selective sensors are available in E-tongue which is also known as "a multisensory system," and advanced mathematical technique is being used to process the signal based on pattern recognition Table 5 Application of ANFIS in the food industry

Applications
Outcomes References Fish oil (i) A model was developed to estimate the oxidation parameters using three different algorithms which are, ANFIS, multilinear perceptron, and multiple linear regression, and it was found that ANFIS model had the best accuracy in predicting the parameters Asnaashari et al. [186] Ice cream (i) The sensory attributes of ice cream were investigated by using the ANFIS model to predict the acceptability of taste with respect to the input parameters and the model achieved a minimum error of 5.11% and high correlation value of 0.93 Bahram-Parvar et al. [187] Indian sweets (Pantoa) (i) The prediction of the heat transfer coefficient during the frying of pantoa using the ANFIS model yielded a high R 2 value of 0.9984, and this prediction is important for designing the process equipment as well as saving energy in commercial production Neethu et al. [188] Orange (i) The developed ANFIS model was able to predict the orange taste and has higher performance when compared to multiple regression model Mokarram et al. [189] Quince fruits (i) The ANFIS model was able to predict the moisture ratio, energy utilization, energy utilization ratio, exergy loss, and exergy efficiency of quince fruit during the drying process with high accuracy of with Rapeseed oil (i) The developed ANFIS could predict the different outputs of rapeseed oil process by oil extraction and cooking at industrial scale, and the model achieved a high correlation coefficient which is around 0.99 Farzaneh et al. [152,153] Salmonella enteritidis (i) Prediction of the inactivation of Salmonella enteritidis by ultrasound was able to be done by the developed ANFIS model with a good accuracy where the correlation coefficient obtained was 0.974 (ii) This study was known to be important in the food industry as the bacteria can cause food poisoning if proper detection is not being done Soleimanzadeh et al. [191] Taro (i) The optimization of extraction conditions of antioxidants from the taro flour can be done by using the developed ANFIS model coupled with response surface methodology (ii) The prediction values obtained from the developed model were validated by comparing with the experimental values, and the results were almost consistent with prediction values from the developed model Kumar & Sharma [192] Vegetables (cantaloupe, garlic, potatoes) (i) The developed model by using the ANFIS system was able to predict the effective moisture diffusivity, specific energy consumption, moisture ratio, and drying rate of the vegetables with a high regression coefficient of 0.9990, 0.9917, 0.9974, and 0.9901, respectively, with minimum error value (ii) Comparison between the ANFIS model and ANN model was carried out, and the results showed that the ANFIS model possess a higher efficiency than that of ANN model Kaveh et al. [193] Virgin olive oil (i) The ANFIS model was able to predict the quality of virgin olive oil samples with high accuracy where the coefficient determination obtained was greater than 0.998 (ii) It was also able to visualize the effects of temperature, time, total polyphenol, fatty acid profile, and tocopherol on the oxidative stability of virgin olive oil Arabameri et al. [194] Yam (i) The prediction of the yam moisture ratio during the drying process showed a high R 2 value with 0.98226 by using the developed ANFIS model Ojediran et al. [195] (PARC) and multivariate data analysis [102]. For example, different types of chemical substances in the liquid phase samples can be segregated using E-tongue. About seven sensors of electronic instruments are equipped in E-tongue, which enabled it to identify the organic and inorganic compounds. A unique fingerprint is formed from the combination of all sensors that has a spectrum of reactions that differ from one another. The statistical software of E-tongue enables the recognition and the perception of the taste. E-tongue comprises three elements specifically the sample-dispensing chamber or automatic sample dispenser, an array of sensors of different selectivity, and image recognition system for data processing (Ekezie, 2015). Samples in liquid forms could be analyzed directly without any preparation while the samples in solid forms have to undergo preliminary dissolution before measurement is carried out. The process of E-tongue system is shown in Fig. 7 below. The ability to sense any taste like a human olfactory system makes it one of the important devices in the food industry, especially for quality control and assurance of food and beverages [103].
In addition, E-tongue has been used to identify the aging of flavor in beverages [104], identify the umami taste in the mushrooms [105], and assess the bitterness of drinks or dissolved compounds [102]. Jiang et al. performed a summarized review on the application of e-nose in the sensory and safety index detection of foods [106]. Moreover, the demand of E-tongue in the food industry market has risen due to the awareness on delivering safe and higher-quality products. The details of recent applications of E-tongue in the food industry are shown in Table 7. The computer vision system (CVS) is a branch of AI that combines the image processing and pattern recognition techniques. It is a non-destructive method that allows the examination and extraction of image's features to facilitate and design the classification pattern [107]. It is also recognized as a useful tool in extracting the external feature measurement such as the size, shape, color, and defects. In general, it comprised a digital camera, a lighting system, and a software to process the images and carry out the analysis [108]. The system can be divided into two types which are 2D and 3D versions. Its usage is not restricted to various applications in food industries such as evaluating the stages of ripeness in apples [107], predicting the color attributes of the pork loin [109], detecting the roasting degree of the coffee [110], evaluating the quality of table grapes [111], and detecting the defects in the pork [112]. The combination of CVS with soft computing techniques has been said as a valuable and important tool in the food industry. This is because the combination of these systems offers good advantages such as an accurate prediction in a fast manner can be achieved. Table 8 shows the combination of CVS and soft computing that has been used in the food industry. Figure 8 shows the working principle of CVS. An example on the utilization of CVS for the quality control is shown in Fig. 9 [113].
Near infrared spectroscopy (NIRS) is another technique in the food industry as there is no usage of chemicals and results can be obtained accurately as well as precisely within minutes or even continuously [114]. In addition, it is known to be nondestructive, cost effective, quick, and straightforward which makes it a good alternative for the traditional techniques which are expensive and labor intensive and consumes a lot of time [115]. The chemical-free method by NIRS makes it suitable to be used as a sustainable alternative since it will not endanger the environment or the human health. It has a wide range of quantitative and qualitative analysis of gases, materials, slurries, powders, and solid materials. Furthermore, samples are not required to be grounded when light passes through it and certain features or characteristics that are unique to the class of the sample are revealed by the spectra of the light. Complex physical and chemical information on the vibrational of molecular bonds such as C-H, N-H, and O-H groups and N-O, C-N, C-O, and C-C groups in organic materials can be provided by the spectra which can be recorded in reflection, interactance, or in transmission modes [114].
The basic working principle for NIRS is shown in Fig. 10. Recently, NIRS has become an interest in food industries to inspect food quality, controlling the objective of the study and evaluating the safety of the food [114,[116][117][118][119]. Several researchers have applied the NIRS in food to obtain its properties for multiple reasons including determining the fatty acid profile of the milk as well as fat groups in goat milk [120]. Apart from that, it is able to aid in the prediction of salted meat composition at different temperatures [121] and in the prediction of sodium contents in processed meat products [122]. The detection and grading of the wooden breast syndrome in chicken fillet in the process line was also able to be performed by using the NIRS technique [123]. Not only that, it is proven to be efficient in determining the maturity of the avocado based on their oil content [124], predicting the acrylamide content in French-fried potato and in the potato flour model system [125], and determining the composition of fatty acid in lamb [126]. There has been a review conducted on the application  Thazin et al. [198] Chicken meat To classify the fresh and freeze-thawed chicken meat Mirzaee-Ghaleh et al. [199] Cow ghee To detect the adulteration of the margarine in cow ghee ANN (i) The ANN model was able to analyze the data obtained from the e-nose with high accuracy Ayari et al. [200] Edible oil To detect the adulteration in oxidized and non-oxidized edible oil ANN (i) The developed ANN model with e-nose was able to detect the adulteration in the edible oil with high accuracy (ii) The classification of the system was compared with other methods and it was given that ANN had the highest classification rate with 97.3% Karami et al. [201] Fish To identify and classify the fish spoilage ANN, PCA (i) The developed model using the PCA and ANN was able to classify the fish according to their spoilage group with an accuracy of 96.87% Vajdi et al. [202]  Wheat grain To determine the granary weevil infestation in stored wheat grains FL (i) The most responsive sensors and specific VOCs generated by insect-infected wheat grains were able to be screened out by the e-noses sensor associated with the fuzzy logic analysis (ii) E-nose was proven to be a potential method for accurate and rapid in monitoring the infestation in stored wheat grain. It is also a reliable method for industries to determine the quality of the product throughout the storage period Mishra et al. [219] of the ANN combined with the near-infrared spectroscopy for the detection and authenticity of the food [127]. The ability of the NIRS system in detecting the physical and chemical properties coupled with soft computing techniques such as ANN, FL, and ML allows the classification and prediction of the samples to be performed rapidly and accurately. Table 9 shows the application of NIRS coupled with AI techniques in the food industry.

Summary on the Application of AI in the Food Industry
From the review so far, it can be shown that AI has been used for various reasons in food industries such as for detection, safety, prediction, control tool, quality analysis, and classification purposes. Ranking of sensory attributes in the foods can be done easily by using the FL model. Not only that, fuzzy logic can be used for classification, control, and non-linear food modeling in the food industry. ES is widely used in the food industry for decision-making process. On the other hand, ANN model is applied widely in the food industry for prediction, classification, and control task as well as for food processing and technology. The supervised ANN method has the ability to learn from examples which allows for the prediction process to be done accurately. Meanwhile, the unsupervised method of ANN is found to be more common for the classification task. Another method that has been utilized for the prediction and classification of the food samples is by using the machine learning (ML) method. ML can be used in solving complicated tasks which involves a huge amount of data and variables but does not have pre-existing equations or formula. This method is known to be useful when the rules are too complex and constantly changing or when the data keep changing and require adaptation. Furthermore, the adaptive neuro fuzzy inference system (ANFIS) is another hybrid AI method that can be used to solve sophisticated and practical problems in the food industry. However, decent data are required for the model to learn in order to perform well. In addition to that, this model is useful for solving analytical mathematical models in the food industry such as studies involving mass and heat transfer coefficients. ANFIS is recommended to be used when complex systems where time-varying processes or complex functional relationships and multivariable are involved. Apart from that, it can be used in descriptive sensory evaluation. These AI algorithms can be combined with other sensors such as the electronic nose, electronic tongue, computer vision system, and near infrared spectroscopy to glean the data from the samples. Both the E-nose and E-tongue have shown to enhance the quality characteristics in comparison to the traditional detection approach [128]. E-nose can be used to sense the odors or gases while the E-tongue can be applied for the identification of the organic and inorganic compounds. Studies involving the examination and drawing out the features of the samples like shape, color, defects, and size can be carried out by using the CVS sensors. NIRS can be utilized to determine the properties or contents in the samples. The data obtained from these sensors is then merged with the AI algorithms and utilizing their computing strengths to accomplish the desired studies.

Advantages and Disadvantages of AI
AI has been used widely in the industry as it offers a lot of advantages compared to the traditional method. All the algorithms are known to be accurate and reliable, but careful selection should be made by considering the advantages and limitations of the algorithms. The different algorithms have their own strengths and weakness, hence choosing them for a particular application in the food industry needs to be looked on a case-to-case basis. The guideline to choose the most appropriate method is given in the next section. The benefits and constraints  To monitor the salt processing of hams salted differently with different formulations Simplified fuzzy ARTMAP neural network (i) The data obtained from the e-tongue was able to be analyzed and classified using ANN (ii) The data was classified using two processing variables which are the processing time and salt formulation (iii) Optimum parameters value for SFAM neural networks were drawn out to be used in the microcontroller device Gil-Sánchez et al. [220] Honey To differentiate different types of honey according to their antioxidant level Fuzzy ARTMAP neural network (FAM) (i) The proposed E-tongue system was able to differentiate different types of honey as well as their total antioxidant capacity level (ii) The ANN fuzzy art map type analysis had a high classification success rate of 100% which indicates that it is a good measurement system Marisol et al. [221] Liquor To classify different types of Chinese liquor flavor using e-tongue with fuzzy evaluation and prediction by SVM SVM & fuzzy evaluation (i) E-nose with the SVM system was able to classify four different flavors of liquor with an accuracy of 100% (ii) The developed system is able to discriminate the samples accurately and the output evaluation language in line with the human perception Jingjing et al. [222] Milk To detect the adulteration of raw milk SVM (i) The developed model was able to determine the adulteration in the samples with a high accuracy values which are all greater than 87% for different types of adulterants in the milk Tohidi et al. [223] Peanut meal To assess the taste attributes of peanut and compare the predictive abilities of the methods used ANN, partial least square (PLS) (i) Good stability and repeatability with respect to the measured signals were exhibited by the sensors in the E-tongue (ii) Different concentrations with the same taste (five types of taste) were able to be discriminated by the E-tongue (iii) RBFNN has a better prediction ability with lower error and higher correlation coefficients than those of the PLS method Wang et al. [224,225] Pineapple To classify the pineapples according to their sweetness level and determine the best algorithm SVM, KNN, ANN, RF (i) Different machine learning algorithms were employed in determining the sweetness of the pineapple, and the best algorithm obtained was the KNN method where it achieved an accuracy of 0.820 (ii) The developed model will be beneficial in industry when the selection of pineapples in large quantities is required Hasan et al. [226]  To discriminate and predict the solid foods as well as to provide an assessment tool for food industries ANN (i) RBFNN was able to distinguish different types of rice with 95% accuracy in classification (ii) Voltametric E-tongue is useful for the qualitative analysis for rice Wang et al. [224,225] Rice To develop similarity analysis combined with artificial neural networks (SA-ANN) in e-tongue for the prediction of rice sensory quality ANN (i) SA-ANN in E-tongue was able to predict the rice sensory quality and carry out systematic analysis (ii) Comparison was carried out between PCA-ANN and SA-ANN, and it was found that SA-ANN has better precision and accuracy compared to PCA-ANN (iii) SA-ANN is a less-labor intensive, quicker method and has potential for rapid and big scale prediction of rice sensory property Lu et al. [227] Sugarcane To characterize and apply voltametric e-tongue for the analysis of glucose from the sugarcane ANN (i) Multilayer ANN with wavelet information was able to process complex responses from the E-tongue (ii) The proposed model is suitable to be used for hydrolyzed samples from sugarcane bases De Sá et al. [228] Tangerine peel To classify tangerine peel of different ages BPNN, ELM (i) The model was able to classify the tangerine peel samples of different ages (ii) Comparison was done for few linear models and non-linear models, and it was obtained that nonlinear models exhibited better performance than linear models (iii) ELM was the best for the classification of the samples with high accuracy followed by BPNN Shi et al. [229,230] Teas To distinguish different types of teas ANN, SVM (i) Different types of teas were able to be distinguished by using the developed system and the compositions of the tea also could be identified Huang et al. [216,231] Tilapia fillets To predict the changes in freshness of tilapia fillets at different temperatures using the combined techniques ANN-PCA (i) E-tongue is able to distinguish the extracts of tilapia fillets stored at different days and different temperatures (ii) The model set up is able to predict the freshness of tilapia fillets stored at different temperatures ranging from 0 to 10 °C Shi et al. [229,230]  Bell pepper To describe the ripeness level of bell pepper automatically ANN, FL (i) An artificial vision system was able to be developed by using the CVS and ANN/FL in predicting the maturity of the bell pepper (ii) The model using RBF-ANN has a higher classification accuracy compared to FL where the maximum accuracy obtained by both models are 100% and 88%, respectively Villaseñor-Aguilar et al. [237] Cape gooseberry To classify the ripeness of cape gooseberry ANN, SVM, DT, KNN (i) All the models were able to classify the ripeness of the cape gooseberry with a high accuracy where the accuracy obtained by all the models was greater than 86% using different color spaces, which indicates that it is a good classifier system Castro et al. [238]  Gluten-free cake To develop a system for quality control of celiacfriendly products FL (i) The developed system was able to study the texture of the cake when different amounts of materials were added to it, and the optimal ingredients value suitable for the gluten-free cake were able to be determined Rezagholi & Hesarinejad [243] Lime To predict the weight of Indian lime fruits ANFIS (i) Different clustering methods were fused with the ANFIS model to improve the accuracy in the classification system, and it was found that the fuzzy C-means clustering (FCM) was the best in predicting the weight of the sweet lime (ii) The developed system was able to predict the weight of the Indian sweet lime fruits accurately Phate et al. [244]

Mushrooms
To determine the appearance quality of mushrooms ANN, FL (i) The accuracy obtained by the image processing system was 95.6% (ii) The artificial neural network was able to determine the weight of the mushrooms, and the fuzzy logic used the data from the CVS and was able to determine the quality of the mushroom Nadim et al. [246] Passion fruits To classify the passion fruits based on their ripeness level Multi-class SVM (MCSVM) The developed system was able to classify the ripeness level passion fruits with an accuracy of 93.3% within 0.94128 s Sidehabi et al. [247] Pork loin To assess the quality of the pork loin according to the industry demand SVM (i) The model was able to predict the quality of the pork loin based on their color and quality attributes based on the industries' demand Sun et al. [248] Potatoes To develop a grading system for potatoes FL (i) Combination of CVS and fuzzy logic allows faster grading of the potatoes as well as reduces the cost required for the manual grading Bhagat & Markande [249] Rice To carry out the qualitative grading of milled rice FL (i) The study was able to conclude that the developed hybrid system can be used in the processing industry for automatic grading of milled rice (ii) Comparison was done between the developed system and experts' judgment, and around 89.80% overall confidence was obtained (iii) The fuzzy system has obtained 89.83% total sensitivity and 97.45% specificity for the quality grading of milled rice

Zareiforoush et al. [250]
Rice To control the performance of rice whitening machines FL (i) The developed automatic control system had an average of 31.3% higher performance speed than that of a normal human operator, and there was an improvement in the quality of the output based on the decision made by the system (ii) The setup flexibility of the system allows alteration to be done according to the preference of each rice mill operator Zareiforoush et al. [55] that each of the algorithm exhibits are explained briefly in Table 10.

Guidelines on Choosing the Appropriate AI Method
Selecting the appropriate algorithm is important when developing the AI model as it can aid the user to attain an accurate, rapid, and cost-saving results. Therefore, a guideline given in Fig. 3 is an important asset prior to achieving best performances in a case study. The primary step in the selection process is that users should define and finalize the objective of using AI in their research or implementation. Prediction, classification, quality control, detection of adulterants, and estimation are among the common objectives of AI applications in the food industries. Next, decision should be made whether sensors such as E-tongue, E-nose, CVS, and NIRS are required to collect the sampling data or not for collecting the data from the samples. Normally, integration with those sensors is conducted to obtain the parameters and characteristics of the samples to be included in the AI algorithms for sample testing purposes. Upon deciding the necessity of the sensors, users should compare and choose the fitting algorithm with respect to their study. Among the most common AI algorithms that have been employed include the FL, ANN, ANFIS, and ML methods. ANFIS has shown to have a higher accuracy, but the complexity of the model makes it less favorable compared to the other algorithms. It is advisable for the users to determine the complexity of the research in selecting the most appropriate algorithm for their studies. Once the selection of the algorithm has been confirmed, the data available are integrated with the AI algorithms. Finally, the testing and validation based on R 2 and MSE are done to analyze the performance of the established model. The AI model has been created successfully once the validation is accepted; otherwise, users should return to the previous step and reselect the algorithm. Figure 11 shows the guideline in choosing and development of the AI model in food industry application.

Trends on the Application of AI in the Food Industry in the Future
The overall trend on the application of AI in the food industry is shown in Fig. 12. From the studies within the past few years, the usage of the AI methods has been observed to increase from 2015 to 2020 and is predicted to rise for the next 10 years based on the current trends. Among the rising factors for the application of AI in the food industry is the introduction of Industrial Revolution 4.0 (IR 4.0). The merging of technologies or intelligent systems into conventional Vegetable seeds To classify the vegetable seeds FL (i) The system can classify the two different types of seeds that look similar which are cauliflower seed and Chinese cabbage seed Garcia et al. [253] industry is what is known as IR 4.0 and can also be called smart factory [129,130]. AI which is categorized under the IR 4.0 technologies focuses on the development of intelligent machines that functions like the humans [131]. IR 4.0 makes a great impact in the product recalls due to the inspections or complains in the food industries. The implementation of the AI integrated in the sensors able to detect the errors during the manufacturing process and rectify the problems efficiently. Apart from that, IR 4.0 also plays a big role in the human behavior as consumers in the twenty-first century often discover information regarding the foods in the internet. The rising concerns on the food quality allow more usage of AI as they are able to enhance the quality of the food and aids during the production process. The highest amount of application of AI in the food industry was seen in the year 2020 as more researchers are carrying out studies using the AI method, and it is believed to continue rising for the upcoming years due to increasing in food demand and the concern on the safety of the foods which are being produced.
The comparison between the AI integration with and without sensors for real-time monitoring in the food industry is displayed in Fig. 13. Integration with external sensors has a higher percentage compared to those without the integration of the sensors in the food industries. The purpose of external sensors was to obtain the data from the samples which are then employed into the AI algorithms to carry out various tasks such as classification, prediction, quality control, and others that have been stated earlier. However, the data collection for the year 2017 showed that the percentage for the AI without the external sensors is greater than that with integration with the sensors. This is due to the high amount of research which was conducted without using the external sensors which are listed in this paper. Based on the

Input
Sensing Device Interpreting Device Output Fig. 9 CVS-based quality control process  Meat To create a system for the detection of meat spoilage AFLS (i) The AFLS model was able to classify the meat into three classes which are fresh, semi-fresh, and spoiled by using the data provided by the FTIR spectrometer (ii) The model achieved a high percentage with a value of 95.94% of correct classification overall which indicates that it can be an effective tool for the detection of meat spoilage Alshejari & Kodogiannis [262] Olive oil To detect the adulteration in olive oils SVM (i) The fusion of NIRS and Raman spectral data could identify the adulterated olive oils effectively and the SVM model was able to predict the dopant contents in olive oil accurately Xu et al. [263] Pears To determine the soluble solid contents in pear ELM (i) The proposed successive projection algorithm and extreme learning machine (SPA-ELM) was able to predict the contents better than the conventional PCA-ELM method Lu et al. [264] Rice To classify the rice according to the compositions and processing parameters RF, PCA, LDA, PLS (i) All the ML techniques were able to identify and classify the rice based on its composition (amylose-based, glycemic index) and the hydrothermal treatment severity with a good performance Rizwana & Hazarika [265] Wheat flour To predict the wheat flour quality using the NIRS Multitarget coupled with SVM, RF (i) The use of multitarget over partial least squares coupled with machine learning algorithm offers more advantage for the parameter prediction from NIRS (ii) The prediction using random forest overcomes the performance of SVM Barbon Junior et al. [266] White asparagus To predict the origin of the asparagus and distinguish the German from imported products SVM (i) The linear SVM could predict the country of origin of white asparagus with an accuracy of 89% and also was able to distinguish the German and non-German products Richter et al. [267] evaluation carried out during this study, it was found that a high amount of research was done on the integration of CVS sensors with the AI methods. It is explainable as CVS sensors are able to provide important parameters such as the shape, size, colors, and defects which are essential for the quality control in the food industry. However, the integration of the system is mainly dependent on the objectives of the researcher and the industrial players and the availability of the data. In short, as the AI world is heading towards 2.0 [132], it can be predicted that the rise in the usage of AI in the food industry is definite and inevitable because of the advantages that they can offer such as saving in terms of time, money, and energy as well as the accuracy in predicting the main factors which are affecting the food industries. Apart from that, in the recent pandemic situation due to the Covid-19 virus, it is predicted that more companies will opt for the usage of AI in their industries to cut down the costs and boost the performance of their company. There have been reports by some of the SMEs that their earnings have dropped and some SMEs have claimed that they could only survive for about 1 to 3 months. The high demand of food and the tight standard operating procedure in the companies during the pandemic situation will encourage the industry players to find an alternative to their problems and AI will be one of them to ensure a smooth operation. (i) The construction and designing of the ES are expensive and rare as it requires expert engineers (ii) Vocabulary utilized by the experts is limited and often is difficult to be understood FL (i) Imprecise, incomplete, and uncertain information can be solved (ii) Simpler and direct results can be obtained (iii) Accountable, noise tolerant, and robust to disturbances (iv) Faster interpretation than ANN and support vector machine method (v) The knowledge base can be extended easily with the extension of the rules (vi) Saves costs and time (vii) Can improve the quality and safety of the products (i) Generalization is not possible as it can only deduce the given rules (ii) Sometimes requires the knowledge of an expert in creating the rules ANN (i) Able to model complex functions or problems accurately and easily (ii) Accurate, robust to disturbances, and noise tolerant (iii) Has the ability to learn from the patterns or examples (iv) Has the generalization ability (v) Affordable, noise tolerant, easy, and flexible method (vi) Solving non-linear problems are more appropriate by using this method (vii) Useful as prediction, classification, and control tool (i) The performance of the model is hard to be explained compared to the others as it appears as a black box model (ii) Requires more time compared to the FL as the suitable number of layers should be determined (iii) Need sufficient and reliable data ANFIS (i) Able to merge details from various resources (ii) Noise tolerant, accurate, and effective method in solving complex problems (iii) It has a higher performance compared to ANN and FL methods (iv) Possesses the benefits from both ANN and FL method (v) Classification and prediction tasks can be done more conveniently (vi) Able to save time and cost overall in general compared to manual methods (i) The data available should be reliable to avoid any confusion or misinformation during the training process as it will affect the final results

Fig. 11
Flowchart for developing AI model

Conclusion and Future Outlook
In conclusion, AI has been playing a major role in the food industry for various intents such as for modeling, prediction, control tool, food drying, sensory evaluation, quality control, and solving complex problems in the food processing. Apart from that, AI is able to enhance the business strategies due to its ability in conducting the sales prediction and allowing the yield increment. AI is recognized widely due to its simplicity, accuracy, and cost-saving method in the food industry. The applications of AI, its advantages, and limitations as well as the integration of the algorithms with different sensors such as E-nose and E-tongue in the food industry are critically summarized. Moreover, a guideline has been proposed as a step-by-step procedure in developing the appropriate algorithm prior to using the AI model in the food industry-related field, all of which will aid and encourage researchers and industrial players to venture into the current technology that has been proven to provide better outcome.
Funding The authors were supported by the Universiti Kebangsaan Malaysia under grant GUP-2019-012.

Conflict of Interest
The authors declare no competing interests.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.