Appearance Defect Detection and Localisation using a Lightweight CNN-based Detector

Automatic visual inspection plays a crucial role in many industrial sectors to assist human operators. This paper studies the general problem of automatic detection of appearance defects with application on wheels surface quality control. An original method is proposed combining image processing and deep learning. This method exploits geometrical knowledge of the manufactured product, which allows splitting the image into homogeneous zones, over which a dedicated lightweight deep learning network is trained to detect and locate anomalies with the highest accuracy. Additionally, the present paper also addresses the issue of training a supervised AI architecture with a limited availability and imbalance dataset containing 100, 000 images but only 1, 000 with defects. We show on this dataset that the proposed lightweight CNN can achieve a high detection rate for low false-positive rates, which is the main goal for applications in an operational context.