A comparative study of transfer learning-based methods for inspection of mobile camera modules

We apply three transfer learning methods using the pretrained AlexNet convolutional neural network (CNN) model to detect defects in camera modules. In experiments, the performance of fine-tuning methods using random initial parameters in less than the two last fully connected layers while using predetermined weights as initial parameters for the remaining layers, showed better performance than other methods. We expect that the transfer learning-based CNN can be effectively applied to camera module inspection systems.