Abstract: In order to realize the surface defect detection and three-dimensional location of casting defects, we obtain the image by binocular vision system, obtain the point cloud data by triangulation method, and normal distribution transformation registration method is used to realize the three-dimensional positioning of the workpiece. Multi model cascading is used to solve the surface defect detection under the condition of few samples and unbalanced sample proportion. Combined with the coding and decoding method of unsupervised model, the network model of defect image repair is obtained by positive sample training, and the defect location is obtained by difference analysis between the repaired image and the original image; the supervised model training is carried out by MASK-RCNN model, and the integration is obtained detect the segmentation model and locate the defect location and category directly. In addition, the physical coordinate system of the moving mechanism, the workpiece coordinate system of binocular imaging and the coordinate of the plane image are converted to obtain the conversion relationship of multiple coordinate systems, and the position information of the plane defect is mapped to the space of the three-dimensional workpiece. Experiments show that this system has good effect in workpiece imaging and detection.

Key words: binocular vision, deep learning, defect detection, multi model cascade