畳み込みニューラルネットワークを用いた外観検査における非対称検出手法の開発

抄録

In this paper, the development of CNN-based appearance inspection method named “asymmetric label smoothing method”, which reduce missing of defective products in appearance is reported. In the manufacturing site, it is necessary to inspect finished products and to remove defective products for quality assurance. The most important function required in appearance inspection is to minimize the number of defective products mistaken for normal ones. The proposed method proactively detects products with suspected defects, thereby reducing the number of missed defective products and over-detected of normal products. In the proposed method, asymmetry is given to normal and defective products, and the label smoothing method is applied among the classes of defective products. Using the model trained by the proposed method, we conducted an experiment of appearance defect detection in automobile clutch discs and succeeded in reducing the number of missed and over-detected compared to the conventional method. The proposed method is expected to be an effective method to improve the accuracy and to reduce the number of missed in defect detection at actual mass production processes.