Abstract
The rise of deep learning, especially in the realm of computer vision, paves ways of leveraging automatic optical inspection systems to a higher level. Convolutional neural networks and its derivatives might be the most widely used architectures for defect inspection tasks. In real cases the amount of collected data is often not large, so transferring learning and data augmentation are necessary. In this paper, we explain some details how we implement the deep learning based AOI system where fully connected layers are replaced by convolutional layers, then a classification heat map is output after post-processing. We examine the performance of our model with two data sets collected in industrial manufacturing cases. We further propose an idea to transfer models pretrained on augmented data of different sizes cropped from original image to the present classification task for possible improvements of the performance.