Convolutional Neural Network Based Surface Inspection System for Non-patterned Welding Defects

In this paper, we propose a convolutional neural network (CNN) based method that inspects non-patterned welding defects (craters, pores, foreign substances and fissures) on the surface of the engine transmission using a single RGB camera. The proposed method consists of two steps: first, extracting the welding area to be inspected from the captured image, and then determining whether the extracted area includes defects. In the first step, to extract the welding area from the captured image, a CNN based approach is proposed to detect a center of the engine transmission in the image. In the second stage, the extracted area is identified by another CNN as defective or non-defective. To train the second stage CNN stably, we propose a class-specific batch sampling method. With our sampling method, biased learning caused by data imbalance (number of collected defective images is much less than that of non-defective images) is effectively prevented. We evaluated our system with a large amount of samples (about 32,000 images) collected manually from the production line, and our system shows a remarkable performance in all experiments.