Introduction

Photos sometimes include unwanted regions such as a person walking in front of a filming target or a trash can on a beautiful beach. Image inpainting is a technique to automatically remove such areas (“damaged regions” in this paper) and restore them [4, 5, 7, 8, 12, 15, 17, 34]. However, it is known that inpainted results vary largely with the method used and the parameters set.Footnote 1 For example, He et al.’s method effectively repairs images including horizontally or vertically repeated textures [12] and Huang et al.’s method is especially efficient for structural images [15]. In the conventional approach to obtain the best inpainting results, a user selects the inpainting technique and tunes the parameters by trial and error, while observing the inpainted results. Since this is time consuming and requires special knowledge, a method for automatically selecting the best result is required.

Unfortunately, no method has been established to determine the inpainting method and its parameters before conducting inpainting. Instead, we feel that methods for image quality assessment (IQA) that assess the quality of inpainted images have possibility to tackle this purpose. Assessing quality of inpainted images is widely acknowledged as a task that is difficult for automation because its judgment is quite subjective. To handle such tasks, many IQA methods for inpainting focus on how the gaze gathers when a human watches an unnatural image [3, 10, 23, 26, 28, 29, 31]. Through the assumption that unnatural removal of unwanted region gathers human attention, they tried to find a way to represent subjective quality by means of objectively measurable indicators. To obtain human attention, most of these IQA methods use a computational visual saliency map, which simulates human gaze density [3, 10, 23, 26, 29].

Although the basic idea of using human attention is reasonable, these existing methods are difficult to apply for comparing the qualities of two inpainted images due to the following two factors. One is the difficulty in estimating human attention. Actual human attention changes by contexts such as the reason for viewing. Isogawa et al. [16] revealed that the human gaze pattern while watching inpainted images is different from any computational saliency maps. The other one is the resolution of a saliency map, which is generally coarse. Thus, it is difficult to apply this method to the current task in which the difference resides in a locally ubiquitous way.

Estimation of subjective quality is not unique to inpainting image. In the research field of subjective-evaluation-estimation, learning-to-rank approaches have been investigated actively [1, 6, 11, 18,19,20,21, 33]. Although for learning based approaches, large and representative training data sets are essential to improve estimation accuracy, accumulating training data is difficult in view of annotation cost and fluctuation of user annotation. The learning-to-rank framework replaces the subjective evaluation tasks as an ordering task without estimating an absolute score. It is considered that it opens up a new era within the research field of subjective evaluation. The problem setting is quite reasonable and reduces preparation costs since it enables learning without absolute scores annotated by human subjects; selecting the better one is rather easier than providing the scores for images to be subjectively evaluated such as inpainted images. We consider that the learning-to-rank framework has the potential to further reduce the training data accumulation by making good use of ordering traits. Recently, some studies have automatically generated and/or augmented training data by image processing [9, 24, 25]. We believe this concept is also applicable to IQA for inpainting.

In this paper, we show how we tackle the task of obtaining the best inpainted result among inpainted images obtained through various methods and parameters. We also propose a new learning-to-rank based ordering approach for inpainted images. Unlike existing IQA methods for inpainting, our method has a new feature that does not use a computational visual saliency map but uses our investigation of human gaze while watching inpainted images. Another important proposal is automatic generation of training data. By making good use of pair-wise learning, we propose automatic generation of training pairs to improve estimation accuracy. The contributions of this paper are as follows:

  • This is the first trial for applying learning-to-rank for IQA of inpainted images.

  • The proposed method enables automatically generated training data to be introduced by making good use of a ranking mechanism, although the learning target is quite subjective.

  • It proposes new image features dedicated to inpainted image quality assessment on the basis of gaze measurement experiments.

This paper is based on our previous conference proceedings [16] and adds a comprehensive investigation on how the proposed features work and a novel method for accumulating training data automatically to improve estimation accuracy. The rest of this paper is organized as follows. In Section 2 we briefly review related work. Section 3 investigates actual human gazes to design effective image features for learning. Section 4 describes the learning based ranking method we propose, which was developed with the knowledge detailed in Section 3. In Section 5 we verify the method’s effectiveness by comparing it with existing IQA methods. We also describe the effectiveness of introducing an auto-generated training set. In Section 6, we conclude the paper with a summary of key points and mention future work to be done.

Related work

This section reviews previous studies. Section 2.1 introduces IQA methods for image inpainting whose purposes are the same as ours. Then, in 2.2, we introduce a learning-to-rank approach that has attracted attention as a method for estimating subjective evaluations.

IQA methods for image inpainting

Estimating quality is one of the difficult issues for image inpainting. The main reasons are the ambiguity in subjective evaluations, and the cost for obtaining training data. Because of the former issue, although many effective IQA methods for degraded images, e.g., burred, compressed, or noised images have been proposed [11, 19,20,21, 32], these methods cannot be applied assessing inpainted images.

To overcome the former issue and to obtain subjective evaluations stably, there are three main approaches: reflecting human reactions such as gaze transition [23, 26, 28], asking subjects to provide their judgments [10], and combining them [29].

The basic concept of human reaction based IQA is that less natural inpainted regions will gather more gazes because of the unnaturalness for human perception. Thus, this method estimates inpainted image quality on the basis of gaze density before and after images are inpainted [28]. To reduce the cost for measuring actual human gazes, many metrics use computational visual saliency maps instead of actual gazes [23, 26, 29]. A computational visual saliency map (“saliency map” for short), is a topographically arranged map that represents estimated visual saliency only from the image. If saliency maps well reflect actual human gaze patterns, substituting them for actual gazes will work well. However, the accuracy of saliency maps is unfortunately quite limited as we mention in Section 3, and thus the performance of saliency map-based IQA methods is also limited.

Learning based approaches that depend on support vector regression (SVR) have also been reported [10, 29]. For these methods, subjectively annotated rating scores are essential for training regression models. Because of the need for absolutely subjective scores, all of the training data should be manually annotated, which is the second issue for building an IQA method for inpainting. Thus generating training sets requires quite high annotation cost. To overcome this issue, our method generates training data automatically as described in Section 4.2.

Ranking based image evaluation for subjective judgment

In many subjective evaluation tasks, it is difficult to provide absolute scores. For example, scoring the degree of smiles is a quite difficult task and the scores may vary largely by question. Since estimating such varied subjective scores is quite difficult, sidestep methods have been widely considered. Learning-to-rank based approaches are now acknowledged as a promising solution. Rather than absolute scores, they provide a learning framework for merely ordering scores among target samples. Coming back to the above cited example, sorting the images by degree of smile is easier than giving smile scores to each image.

Among learning-to-rank approach variants, pairwise learning-to-rank methods have gathered attention due to the ease with which they can be implemented. They have been frequently applied for estimating preference order [1, 6, 18, 33]. Chang et al. estimated the age of a single face image [6]. Yan et al. obtained the most visually appealing color enhancement of an image [33]. Abe et al. estimated the surface qualities of an object, such as glossiness or transparency from its images [1], and Khosla et al. estimated the most memorable region inside images [18].

Learning-to-rank has also been introduced in IQA methods [11, 19,20,21]. Gao et al. [11] and Xu et al. [19, 20] proposed blind image quality assessment frameworks for degraded images, e.g., blurred or compressed images or images with white noise. In addition, Ma et al. introduced learning-to-rank to assessing retargeted images [21]. Although retargeted images are quite different from general degraded images that existing IQA methods deal with, they examined and investigated the effects of learning-to-rank based IQA for image retargeting.

Unlike existing methods, the method discussed in this paper focuses on assessing inpainted images. Since estimating the quality of inpainted images is a quite different task than assessing other deteriorated images, we designed new image features dedicated for assessing inpainted images. In addition, this paper shows that by using pairwise learning traits we can produce training data automatically and use the data to improve estimation accuracy.

Toward effective image features: eye gaze investigation

Many IQA methods use visual saliency maps as substitutes for actual gazes. However, we have doubts about the coherence of computational visual saliency and actual human gazes, especially when observing inpainted images. Therefore, before we go into the proposed method, we will describe an eye gaze measurement experiment we conducted for two purposes. The first was to show the difference between measured gazes and the saliency map and to reveal the difficulty in using saliency maps instead of actual human gazes for IQA. The second was to analyze the region and features within inpainted images we should focus on to assess the quality of the images on the basis of measured gazes and the corresponding subjective evaluations.

Procedure and set-up of eye tracking experiment

We conducted this experiment with the aim of verifying coherence between the measured gazes and the saliency map. We also obtained subjective scores for each image to investigate how gazes affected the total subjective image quality. Figure 1 shows the test procedure, in which subjects repeated three tasks: (a) stare at a white cross on a black background for two seconds to fix their initial viewpoint, (b) observe images for 10 seconds, and (c) provide 5-point opinion scores representing image quality unnaturalness. The scores 1-5 respectively corresponded to Very noticeable, Rather noticeable, Slightly noticeable, Hardly noticeable, and Unnoticeable. Higher scores are better since they indicate that the unnaturalness that occurs with inpainting is unnoticeable.

Fig. 1
figure 1

Three step test procedure of the preliminary experiment conducted to elucidate the relationship between human gaze and subjective scores. Subjects are required to; a stare at the white cross to fix initial viewpoint, b observe an image, and c provide a 5-point opinion score (subjective assessment) of the image

The observed image in task (b) includes original images and inpainted images. Original images are images that are not inpainted. Inpainted images were generated with two methods, i.e., those reported by He et al. [12] and Huang et al. [15]. The subjects had no prior knowledge on the types of images displayed (i.e., whether they were original images or images that had been inpainted by using the methods reported by He et al. [12] and Huang et al. [15]). To prevent the subjects from having prior knowledge of the material, we generated three types of images generated from each of 100 original images. We asked 24 subjects (8 males and 16 females) with normal vision to report the image quality after observing the displayed images. These subjects were divided into three groups and the subjects in each group watched the same type of image. Each subject watched 100 images. We applied a stationary Tobii eye tracker for gaze measurement. The LCD monitor used for stimulus presentation was 21 inches (1280 × 1080 pixels). The monitor-observer distance was 60 cm.

Integrity between computational saliency maps and human visual attention

In Fig. 2, (a) shows an inpainting target image and (b) shows the inpainted result in which an undesired man standing in front of a boat was removed. Measured human attention is overlaid in (c). Calculated visual saliency maps obtained with calculation methods proposed by Hou et al. [14], Achanta et al.’s [2], and Walther et al.’s [30] are respectively shown in (d), (e), and (f). These maps have actually been used for assessing the image quality of inpainted images; they were respectively used by Voronin et al. [29], Trung et al. [26], and Oncu et al. [23]. From these maps, we can observe that the resolution of computational visual saliency maps is quite coarse and their results are significantly varied. Additionally, saliency maps are quite different from human visual attention. As shown in Fig. 2b, inpainting failed to fill the shape of the boat. Because this failure produces significant unnaturalness, the most salient areas for actual human gazes were those around the damaged boat (See Fig. 2c). The areas around the boat in (d) and (e) were somewhat salient, but were more salient in other areas (e.g., the other boats or the oars). In (f) no saliency around the boat was represented at all. These results suggest that it is difficult to use computational visual saliency maps as a substitute for human gazes. Thus, it is essential to come up with new image features that represent such unnaturalness.

Fig. 2
figure 2

Comparison between observed human visual attention and computational visual saliency. a Inpainting target image. b Inpainted image. c Human visual attention overlaid on (b) (red gathers more gazes). ef Computational visual saliency overlaid on (b) (red gathers more gazes). Saliency maps are (d) Hou et al.’s [14] used in Voronin et al.’s metric[29], e Achanta et al.’s [2] used in Trung et al.’s metric [26], and (f) Walther et al.’s [30] used in Oncu et al.’s metric [23]

Correlation between subjective quality and human visual attention to damaged region contours

Human gazes are potentially an excellent means for assessing inpainting quality, and a metric based on human gazes was proposed by Venkatesh et al. [28]. This metric categorizes eye gaze position into two categories, i.e., inside and outside damaged regions. They use the difference of amount of gazes between pre- and post- inpainted images. We were inspired by this simple and effective idea and so tackled further analysis of eye gaze patterns in categories other than inside and outside damaged regions. This section shows how we analyzed what we should focus on to assess the quality of the inpainted images on the basis of knowledge of human attention and corresponding subjective evaluations. We believe that this knowledge will be useful in developing an IQA method for image inpainting.

We analyzed the characteristics of observed gaze and corresponding MOS levels. The MOS values are the average of the 5-point annotated scores provided by subjects as described in the previous section. We first investigated on where human beings tend to watch for inpainted results. Fig. 3a shows an example gaze histogram for the inpainted image in Fig. 2b. Its vertical axis is the time the gaze was oriented and the horizontal axis is the distance from the contour of a damaged region, where a negative value means inside the damaged region. As shown in the histogram, around the contour, i.e., distance = 0, gathers more gazes, which indicates the contour of the damaged region tends to be salient.

Fig. 3
figure 3

Gaze measurement results; a gaze histogram with Fig. 2b, b relationship between the subjective score and density of gaze within the vicinity of damaged region’s contour

To be more specific, Fig. 3b shows the relationship between subjective score and gaze density within the vicinity of a damaged region’s contour for two different inpainting methods. Green and red points are for the methods proposed by He et al. [12] and Huang et al. [15]. Here, we set the contour vicinity to be within 30 pixels from the contour. This corresponds to the 1.0 degree view angle we used in our experimental setup. As shown in Fig. 3b, the correlation coefficients r for the inpainted results provided by He et al. and Huang et al. are respectively r = − 0.63 and r = − 0.47. These results indicate a high negative correlation exists between subjective scores and gaze density around the contour. Thus there is a high probability that the image features around the contour are important for preference estimation.

Proposed method

Now we are ready to describe our proposed method. Our goal is to obtain the best result among inpainted images that are generated with different inpainting methods and parameters. In Section 4.1, we describe our pairwise ordering method and in Section 4.2 show how we automatically generated large auto-generated training set in which manual intervention was not required. Then, in Section 4.3 we show effective image features for this approach, which include the knowledge given in Section 3.

Ranking by assessing image quality with learning

Aiming at tackling the difficulty in reflecting subjective evaluations for inpainted results to scores, we based our preference order estimation on a learning-to-rank approach. Figure 4 shows the overview of our proposed method.

Fig. 4
figure 4

Proposed method overview

Before we explain the details, let us briefly explain a typical pairwise learning-to-rank algorithm. This algorithm premises a ranking function f(x), which computes the strength of the target attribute for each sample, as described below. Hereafter, we use xi to denote a feature vector extracted from sample image zi. The f(x) is trained so that the ordering of the output value from the function f(x) reflects the user annotated preference order zizj between image pairs. In a word, the function f should satisfy the following formula:

$$\begin{array}{@{}rcl@{}} z_{i} \succ z_{j} \iff f(x_{i}) > f(x_{j}) \end{array} $$
(1)

We modeled f with the linear function f(x) = ωx. Then inequalities (1) can be written as below.

$$\begin{array}{@{}rcl@{}} z_{i} \succ z_{j} \iff \omega^{\top}(x_{i} - x_{j}) > 0 \end{array} $$
(2)

This mirrors the problem of binary classification. To implement the formulation that uses binary classification to calculate the preference order of pairs of images, the pair-wise learning to rank approach is widely used. From among the various methods yielding pair-wise learning to rank, we adopted RankingSVM [13] as it is used widely [1, 18, 33] due to its effectiveness and ease of implementation.

In our method, function f is trained with the pair of image feature vectors described in Section 4.3 with their preferences. We call a training data set of this type a “query” (see Fig. 4e). Basically, all preferences for generating queries are manually annotated by subjects. First, inpainted images with several parameters are generated with a masked image as shown in Fig. 4i–a. These results are used to make a pair of train images like those in Fig. 4i–b. Then, as shown in Fig. 4i–c, subjects were asked to provide their preference judgment (x) to a pair of inpainted images (z). Subjects’ preferences are reflected to all image pairs (Fig. 4i–d).

However, the training data shortage problem still remains. To solve this problem, we additionally propose a way to automatically generate training data as shown in Fig. 4ii. In this case, auto-generated images (Fig. 4ii–a) are used instead of inpainted images. The difference is that we already know that preferences depend on the degradation level. Thus we can skip to annotation; images with preferences (Fig. 4ii–d) are directly generated with image pairs (Fig. 4ii–b). In the next section we will describe how we designed a way to automatically generate a training set for which manual annotation is not required.

Auto-generated training data

Existing learning-based IQA methods for inpainting learn the relationship between image features and corresponding scores provided by subjects. Thus, they require user annotated samples. Here, we propose one effective solution, i.e., an IQA method for inpainted images by pairwise ordering. It is effective because it does not need any absolute scores but only pairwise relationships.

We add some distortions, such as proportional changes in pixel values or applying a low pass filter, to the original images that tend to occur as the result of inpainting. Several levels of such distorted images and original images generate training data with the assumption that increased distortion lessens preference. Of course, the original image has better quality than the distorted image. Because our method requires only pairwise relationships, not absolute scores, this simple relationship in which images become more distorted can work as a training data source. Figure 5 shows examples of several levels of auto-generated images for training. The i-th auto-generated train image Ii is synthesized by combining the original image Iorig and the i-th distorted image Di as below.

$$\begin{array}{@{}rcl@{}} I_{i}(x, y) = \left\{\begin{array}{ll} D_{i}^{\gamma}(x, y) & ((x, y) \in {\Omega}) \\ I_{orig}(x, y) & (otherwise) \end{array}\right. \end{array} $$
(3)
$$\begin{array}{@{}rcl@{}} \gamma = \left\{\begin{array}{ll} c & (brighter \, or \, darker \, color \, distortion) \\ b & (blur \, distortion) \end{array}\right. \end{array} $$
(4)

where Ω is a masked region to reflect distortion (see Fig. 5a).

Fig. 5
figure 5

Auto-generated images for training set. a Original image with region to be synthesized, which is masked in red. b Multi-levels of auto-generated images with (i) brighter color distortion, (ii) darker color distortion, (iii) blurred distortion

The reason we apply color and blur distortion is to simulate typical failures occurring as a result of inpainting. Human attention is considered to be quite sensitive to unnaturalness that is produced by differences in brightness and frequency components in images. To represent unnaturalness of this type, we focus on three types of distortion, i.e., that caused by brighter colors, darker colors, and blurring. The former two distortions represent undesired inpainted results in which the colors of inpainted regions are brighter or darker color than those outside the contour of the inpainted region. The latter represents undesired inpainted results in which there is edge discontinuity around the inpainted region. Distorted images of these types are generated as follows.

Brighter and darker color distorted images

We define the i-th brighter/darker damaged image as below:

$$\begin{array}{@{}rcl@{}} D_{i}^{(c)}(x, y) = I_{orig}(x, y) + \alpha \, \beta \, i \end{array} $$
(5)
$$\begin{array}{@{}rcl@{}} \beta = \left\{\begin{array}{ll} + 1 & (brighter \, color \, distortion) \\ -1 & (darker \, color \, distortion) \end{array}\right. \end{array} $$
(6)

where α is a scalar parameter. In the work we report in this paper, we set α = 10.

Blurred distorted image

We define the i-th blurred distorted image as below:

$$\begin{array}{@{}rcl@{}} D_{i}^{(b)}(x, y) = \sum\limits_{(k, l)}I_{orig}(x-k, y-l) \, G(k, l) \end{array} $$
(7)

Here, G is a 2-dimensional Gaussian distribution with kernel size γ and we set γ = 2i + 1.

Features for learning-to-rank

Using the observation provided in Section 3, we designed image features for our framework. We call this image feature patch-based contour consistency (PBCC). As we described in Section 3, human perception is quite sensitive to color or edge discontinuities between in/out of damaged/distorted regions, which we combined to design the PBCC.

The PBCC consists of the following two components: (1) differences between in/out of damaged/distorted region, and (2) normalized image around the contour. The former represents continuity across the contour of the damaged/distorted region. The latter represents the relative quality of images; the coherence of image quality between the inside and outside parts of a damaged region largely affects subjective quality. Thus, even if the image quality within the damaged region is the same, its perceptive quality varies depending on its surrounding region’s quality.

To make the features dedicated for evaluating inpainted images, these components are computed along contours of the damaged/distorted region as shown in Fig. 6. We set the features as x = (Xd,Xs), where Xd and Xs respectively represent the first and the second components. Xd and Xs are computed as below;

$$\begin{array}{@{}rcl@{}} X_{d} &=& || S(P_{in}) - S(P_{out}) ||^{2}_{2} \end{array} $$
(8)
$$\begin{array}{@{}rcl@{}} X_{s} &=& \frac{{\sum}_{p\in\delta{\Omega}} S(P_{out}(p))}{{\sum}_{p\in\delta{\Omega}}1} \end{array} $$
(9)

where Ω and δΩ respectively denote a damaged/distorted region and its contour. Equation (8) represents squared 2-norm. Pin(p) and Pout(p) show masked or non masked regions in patch P(p), which is centered at point p (See Fig. 6). In addition, S(Pin(p)) and S(Pout(p)) represent average features of Pin(p) and Pout(p) as shown below.

$$\begin{array}{@{}rcl@{}} S(P_{in}(p)) &=& \frac{ {\sum}_{q\in P(p) \cap {\Omega}}\,\mathbf{s(q)}}{{\sum}_{q\in P(p) \cap{\Omega}} 1} \end{array} $$
(10)
$$\begin{array}{@{}rcl@{}} S(P_{out}(p)) &=& \frac{ {\sum}_{q\in P(p) \cap \bar{\Omega}}\,\mathbf{s(q)}}{{\sum}_{q\in P(p) \cap \bar{\Omega}}1} \end{array} $$
(11)
Fig. 6
figure 6

Damaged/distorted region and its contour. Pin(p) and Pout(p) show masked or non masked regions in patch P(p), which is centered at point p

To the extent of the work we report in this paper, we used s(p) = (u(p),v(p)), where u(p) = (uR(p),uG(p),uB(p)) and v(p), each denoting RGB pixel values and edge strength.

Experiments

This section describes how we investigated the effectiveness of the proposed method. We will start by detailing the experimental setups used in Section 5.1. We will then verify the effectiveness our method by using the image features we derived and auto-generated training data as described in Sections 5.2 and 5.3, respectively. Section 5.4 then compares our method with the other existing IQA methods. For easy understanding, Fig. 7 shows the flow chart of the experiments conducted in this section.

Fig. 7
figure 7

The flowchart for experiments conducted in Section 5. In Section 5.2, we investigated the efficacy of our proposed image representation. Section 5.3 verifies the effectiveness of auto-generated training data and Section 5.4 compares the estimation accuracy of our method and previous methods to assess their performance

Experimental setup

To generate training and test images, 111 images with manually masked damaged regions were prepared. The 111 images were inpainted with two inpainting methods [12, 15] and six parameters (3 patch sizes and 2 levels of multi-scale parameters) were used for each method. The quality of the images was evaluated by 24 subjects (12 males and 12 females) with normal vision. To make the users’ judgment easy, we randomly displayed a pair of inpainted images side-by-side as shown in Fig. 8. Subjects were asked to choose one of three options: r: right image is better, l: left image is better, and n: no preference order (i.e., it is hard to decide which one is better or which one is worse). Excluding inpainted images with extremely poor quality, we obtained 2,466 image pairs. We excluded poor quality images because they might change subjects’ judgement criteria during the experiment.

Fig. 8
figure 8

Annotation interface for obtaining training data. Two different inpainted results are displayed side by side. Subjects annotate their preferences among three options: r: right image is better, l: left image is better, and n: no preference order

We implemented RankingSVM with SVM Rank [27] with Radial Basis Function (RBF) as the kernel function (γ = 2− 7), and the regularization parameter (C = 2− 5). We used a desktop PC (Intel Core i7, 3.4GHz CPU, 32GB memory) and used Matlab to implement the existing method, and Visual Studio 2012 to implement the proposed method.

Performance comparisons for different image features

This subsection verifies the effectiveness of PBCC, the proposed image feature described in Section 4.3. Table 1 compares the performances attained with seven different image features: Fall [11], GIST [22], EMD kernel [10, 29], Saliency, Fin, Fout, and PBCC. Note that in verifying performance with these image features we used the same training data and estimator; only the image features were different.

Table 1 Performance comparison for different image features [%]

Here, we briefly introduce each of the compared features. Fall was originally used for learning-based-IQA of degraded images. GIST is one of the most commonly used global image features and is used for learning-based IQA for retargeted images. The EMD kernel is calculated on the basis of EMD and has been used in learning-based-IQA for inpainted images. Fin and Fout are the original features of this paper and represent the inside and outside parts of inpainted regions. The two methods represented by (12) and (13). were used for verifying the effectiveness of contour consistency on which PBCC focuses. Saliency is for comparison of computational saliency maps. The same as PBCC, Saliency is calculated so that it represents the inside and outside parts of inpainted regions and depends on how PBCC is claculated. We used Walther et al.’s computational saliency map [30], as it is used by previous work [23].

$$\begin{array}{@{}rcl@{}} F_{in} = \frac{{\sum}_{q \cap {\Omega}}\,\mathbf{s(q)}}{{\sum}_{q \cap {\Omega}} 1} \end{array} $$
(12)
$$\begin{array}{@{}rcl@{}} F_{out} = \frac{{\sum}_{q \cap \bar{\Omega}}\,\mathbf{s(q)}}{{\sum}_{q \cap \bar {\Omega}} 1} \end{array} $$
(13)

Table 1 shows the estimation accuracy obtained, which is the ratio at which the preference order is correctly estimated among annotated pairs. As can be seen from the table, PBCC correctly estimated the image pair preferences at 70.10%, as opposed to the 40.45% to 60.80% obtained with other methods. These results confirm the effectiveness of the proposed PBCC.

Verification of effectiveness depends on the amount of auto-generated training data

In this section we will show the results obtained in an investigation we conducted, which indicate that the system performance changes depending on the size of the auto-generated training set. First, we will show the results obtained with N = 0, in which no auto-generated images are included. Second, we will show how an auto-generated training set affects the performance. We used 100 original images to generate auto-generated image levels of N = 1 to 10. Each level included the three types of distortion described in Section 4.2. N levels of auto-generated images make N+ 1C2 combinations of pair images, including comparisons with the original image. All of these N+ 1C2 pairs are used for training. We investigated the effect of the amount of data with N = 1 to 10. Figure 9 shows .a line graph in which the left y-axis shows estimation accuracy depending on N, which means the number of auto-generated images. Estimation accuracy without any auto-generated training set is annotated with a blue line for reference. The ratio of subjectively annotated training data to the data as a whole is shown by the orange line for the right y-axis. The training set in which N was 3 consists of 900 pairs of samples based on the 100 original images, with three types and three levels (six distortion combinations). In the same way, the training set in which N was 4 consists of 1800 pairs based on the 100 original images, with three types and four levels (10 distortion combinations). These results indicate that a larger training set is more effective; however, a set larger than certain levels of images results in worse performance.

Fig. 9
figure 9

Investigation of performance depending on the amount of auto-generated training set. Estimation accuracy depending on distortion level N is shown with magenta line with left y-axis. Accuracy without any auto-generated data is shown in blue line. Ratio of subjectively annotated training data to whole data is shown in orange bar graph with right y-axis

One of the possible causes for this is that training sets with similar data may result in worse total prediction accuracy. If N is increased, a similar training set generated with the same original images will also increase. This additional data may decrease the effectiveness of learning. We also consider that another cause may be that the rate of subjective annotated data decreases as the amount of auto-generated data becomes larger. In the next section, we will show how we used two settings with different amounts of auto-generated data as our proposed methods: N = 0 with no auto-generated images, and N = 3, which was most effective for learning.

Comparison with existing methods

We conducted experiments in which we compared our method with other IQA methods for image inpainting, i.e., ASV S and DN by Ardis et al. [3], \(\overline {GD_{in}}\) by Venkatesh et al. [28], and BorSal,StructBorSal by Oncu et al. [23]. Because we did not use an actual eye gazes for this experiment, we used a saliency map instead of human gaze for \(\overline {GD_{in}}\); this is the same method that was used in the comparison experiment reported by Oncu et al. [23]. For our method we used N = 0 without an auto-generated training set and N = 3 with such a set; the results obtained were presented in Section 5.3.

Table 2 shows the prediction accuracy obtained for each metric. Our method without/with auto-generated training data correctly estimated the image pair preferences at 70.10% and 73.87% respectively, as opposed to the 50.46% to 60.22% obtained with other metrics. Thus, the improvement our method achieved over existing methods was around 13 percentage points.

Table 2 Prediction accuracy comparison with existing image quality assessment metrics [%]

Figures 10 and 11 respectively show example outputs of correct and incorrect estimation obtained with our method. In both figures, (a) is an original image with a damaged region marked in red, and (b) and (c) are the inpainted results annotated by subjects as (b) ≻ (c). In Fig. 10, (b) images are successfully inpainted while example (c) was an inpainting failure in terms of both color consistency and structure. Our method successfully estimates preferences for such image pairs. We consider that because our image feature design focuses on the unnaturalness produced by color or structural discontinuity, it works well as the Fig. 10 results show. It is especially notable that none of the other existing metrics correctly estimated the preferences for the top left images in Table 2. Other than StructBorSal [23], these methods also failed to estimate preferences for the bottom left images.

Fig. 10
figure 10

Correctly ordered images with proposed method; a original image with damaged region masked in red while b and c are inpainted pairs of images that subjects annotated as (b) ≻ (c)

Fig. 11
figure 11

Incorrectly ordered images with proposed method; a original image with damaged region masked in red while b and c are inpainted pairs of images that subjects annotated as (b) ≻ (c)

We consider that these methods failed due to the uncertainty of the computed visual saliency maps. To demonstrate the cause of the previous method’s failure, Fig. 12 shows a saliency map overlaid on the left half images of Fig. 10. In Fig. 12, (a) to (c) correspond to (a) to (c) in Fig. 10; the original image and inpainted pairs of images. All subjects answered that (b) was better than (c).

Fig. 12
figure 12

To show the cause of the other existing methods’ failure, a saliency map is overlaid on the left top and bottom images in Fig. 10. ac are related to Fig. 10; original image and inpainted images. Upper images show that there are no significant differences between the two inpainted images. In the lower images, b gathered more gazes although subjects preferred (b)

The upper (c) in Fig. 12 includes an inpainting failure around the stairs. However, neither (b) nor (c) gather saliency around the inpainted region. Also, the lower (c) in Fig. 12 has color and structural discontinuity, which generates huge unnaturalness. However (b) gathers more saliency on un-inpainted regions. We consider that this type of uncertainty and instability in saliency maps impede IQA quality.

Our method’s limitation is shown in Fig. 11. In this figure, (b) has a blurred region and also has color discontinuity, but it is relatively natural in context. In (c), if we hide the left half of the picture, it is quite natural. However, structural unnaturalness occurs in the left half of the image and generates context unnaturalness. Because human perception is considered to be more sensitive for contextual failures, subjects preferred (b). Currently our method does not consider any semantic information, thus for such image pairs it generates ordering failures.

Conclusion

In this paper we described an image quality assessment method we developed for image inpainting. Three key ideas of our method are that (1) we use a ranking-by-learning algorithm to estimate the ordering of inpainted images on the basis of subjective quality, (2) our ranking system easily introduces auto-generated training data for more effective learning, and (3) we introducing image features that reflect differences around a contour of damaged regions on the basis of gaze measuring experiments which showed that a high negative correlation exists between subjective quality and gaze density around the contour. Unlike existing image quality assessment (IQA) methods for image inpainting, ours makes it possible to introduce auto-generated training data, due to introduction of our pairwise learning. Preference order estimation experiment results suggest the method’s efficacy. Especially with auto-generated training sets, the estimation performance was about 13 percentage points higher than that of existing IQA methods. In future work, we will introduce other image features such as describing semantic unnaturalness inside inpainted images.