Image dataset creation and networks improvement method based on CAD model and edge operator for object detection in the manufacturing industry

Creating image dataset for object detection is a time-consuming and laborious work, seriously hindering the rapid application of object detection in the industrial manufacturing field. To reduce time and cost of object detection application, a method of image dataset creation and networks improvement based on CAD model and edge extraction operators is proposed. It can quickly generate effective training dataset without any tedious work and make the object detection networks obtain good detection performance. The method consists of three steps: capture-images-automatically, cut-and-paste and networks-improvement. To improve the performance of the detection networks, edge extraction operators are used to obtain the common features of the synthetic images and the real images. These edge extraction operators include Sobel edge, Laplacian edge, Canny edge and adaptive threshold edge, and the experimental results show that the adaptive threshold edge achieves the best effect. In addition, a class-weights is adopted to improve the AP of hard-to-detect parts. Ten mechanical parts of a 3D-printed aero-engine are used to evaluate this method. The results show that the improved networks (yolov5s) trained with the synthetic images can achieve 99.08%, 93.83% and 98.91% of the average recall, average precision and mAP, respectively. Taking into account the time, cost and detection performance, the proposed method is much better than the traditional method and current advanced method. The proposed method is feasible for object detection in many industrial scenarios where CAD models of products can be easily obtained.