부산대학교 도서관

A novel metal surface defect recognition scheme based on improved lightweight YOLOv4 (You Only Look Once version 4) is proposed for workpiece defects recognition in mass production. First, the image preprocessing method accurately cuts the region of interests. Second, use the lightweight network model MobileNetV2 (Mobile Networks Version2) and depthwise separable convolution replace the original backbone feature extraction network and 3 X 3 Standard convolution of YOLOv4 respectively. Therefore the model size is greatly reduced. The detection accuracy of this method for workpiece defects reaches 90.63%, and the detection speed is 33 frames per second. Compared to the original YOLOv4 model, the model size is reduced by 82.1%, the recognition accuracy is increased by 1.8%, and the processing speed increases by 150%.