Surface defect detection of mechanical workpiece based on DSSD model

LI Lan1, XI Shu-shu1, ZHANG Cai-bao1, MA Hong-yang2
(1.School of Information and Control Engineering, Qingdao University of Technology, Qingdao 266000, China;
2.School of Science, Qingdao University of Technological, Qingdao 266000, China)

Abstract: Aiming at the problem of surface defect detection of mechanical workpiece, the types and positions of surface defects of the workpiece were studied, and the object detection algorithm in deep learning was summarized and analyzed. A detection method based on deconvolutional single shot detector （DSSD） model was adopted to detect surface defects of the workpiece. In this method, the defect image of different workpiece and position was acquired by scanning electron microscope ,the data set of surface defect was established, and the data set was expanded. Then the number of network layers of the DSSD network model deconvolution module was simplified to reduce the computational complexity. Finally, the simplified DSSD model was used to train and test the data set. The results show that the detection efficiency of DSSD model is higher than that of YOLO, Faster R-CNN and SSD, and it can detect the surface defects of workpiece more accurately and quickly, providing a new idea for defect detection in actual industrial scenes.
Key words: workpiece defect; deconvolutional single shot detector（DSSD） model; object detection; convolutional neural network