Gear classification for defect detection in vision inspection system using deep convolutional neural networks

For the purposes of accuracy and speed in industrial inspection, many companies heavily depend on human resources rather than automated systems. Nowadays, the most accurate method of image classification is deep learning. As demonstrated in ImageNet challenge, there is still no method that outperforms deep learning. Therefore, for automatic detection of defective gears, we propose the use of deep learning with two kinds of classification approaches, namely the Naïve approach and the fine-grained approach. The Naïve approach allows deep convolutional neural networks (CNN) to directly classify defects and non-defects in gear images, whereas the fine-grained approach harnesses an image processing technique before using CNN. Our experimental results show that there is a tradeoff between these two approaches: the Naïve approach is better in terms of processing time while the fine-grained approach is better in terms of accuracy.