Detection of minute defects using transfer learning-based CNN models

Abstract

In this paper, a design and training tool for convolutional neural networks (CNNs) is introduced, which facilitates to construct transfer learning-based CNNs based on a series-type network such as AlexNet, VGG16 and VGG19 or a directed acyclic graph (DAG)-type network such as GoogleNet, Inception-v3 and IncResNetV2. Minute defect detection systems are developed for resin-molded articles by transfer learning of AlexNet. AlexNet has the shallowest layer structure and the smallest number of weights within the six powerful networks, so that it is selected as the first CNN for evaluation. In the transfer learning process, after the last fully connected layers are replaced according to the number of categories needed for new tasks, an additional fine training is conducted using training images including small typical defects. In experiments, transfer learning-based AlexNet$\_$6 and AlexNet$\_$2 are obtained to deal with six and binary classification tasks, respectively. Then, our originally designed 15 layers CNNs named sssNet$\_$6 and sssNet$\_$2 are also prepared and trained for comparison. Finally, AlexNet$\_$6 and sssNet$\_$6, AlexNet$\_$2 and sssNet$\_$2 are quantitatively compared and evaluated through classification experiments, respectively.