Deep Learning-Based Automatic Detection of Defective Tablets in Pharmaceutical Manufacturing

Abstract

With many tablets produced everyday in manufacturing plants, the pharmaceutical industry needs automatic, highly accurate methods for inspection of tablet quality. Detecting defective tablets is of importance to reduce unqualified products to consumers. In this paper, we propose a deep learning method combining image processing and deep convolutional neural networks (DCNN) for detection of defective tablets using images captured by a multiple-camera inspection system. A dataset of 6000 images of tablets labelled either GOOD or NOT-GOOD were collected at a pharmaceutical factory using commercial camera inspection systems. After collecting and labelling, the images were preprocessed to normalize intensity values. The entire dataset was split into a training set (50%, 3000 images), a validation set (16.6%, 1000 images) and a testing set (33.3%, 2000 images). We trained DCNN ResNets (ResNet50, ResNet101) and DenseNets (DenseNet169, DenseNet201) models on the training set and validated them on the validation set. We applied transfer learning techniques by using pre-trained models that had been trained on the ImageNet dataset in combination with data augmentation and training strategies such as learning rate rescheduling overtime. We compared our deep learning methods with various machine-learning ones such as Support Vector Machine (SVM), K-Nearest-Neighbors (KNN), AdaBoost that used intensity histograms as features. Tuning hyperparameters were performed to seek the best hyper-parameters and algorithms. We achieved best performances using the deep learning models as the ResNet50, and DenseNet169 obtained 96.60% ± 4.9% and 94.13% ± 4.2% accuracies (ACC), respectively. In contrast, SVM achieved 87.75% ACC, KNN achieved 76.09% ± 7.7% ACC while AdaBoost achieved 81.25% ACC, respectively.