CAD Cut-piece Retrieval Method Based on Representation Learning

The accuracy of retrieving the CAD file corresponding to the cut-piece from the CAD library determines the efficiency of the cut-piece automatic detection system and the digital management of the cut-piece inventory. However, the materials of the cut-piece are complex and diverse (including cowhide, artificial leather, fabric, etc.), contain multi-layered shapes and structures, and are prone to deformation. The traditional registration algorithms cannot satisfy the accurate registration between CAD and cut-piece. In order to solve the problem of high-frequency translation and complex shape of the cut-piece, we propose an image retrieval algorithm based on representation learning. This paper takes FaceNet as the basic framework. First, the loss function combining Softmax loss and center loss is used to maintain the discriminant ability between classes and optimize the cohesion of the ternary loss function. Second, the BlurPool module is used to enhance the anti-aliasing of the downsampling process and enhance the translation invariance of the network while maintaining network performance. Finally, the multi-feature deep fusion (FRE) module designed in this paper is used to solve the problem of reduced model expression ability caused by different sizes of cut-piece. A series of experiments on our dataset show that the proposed representation learning network is more accurate in the SVM classifier, and it only takes 30ms to predict a single cut-piece.