Hash Self-Attention End-to-End Network for Sketch-Based 3D Shape Retrieval

To improve the feature clustering of sketches and 3D models,a feature extraction network with self attention and hash regularization is proposed.Firstly,the 3D model is rendered to obtain the different views.Secondly,the self attention layer is embedded in the shared weight network.In this way,the clustering of sketches and views is improved through the auto-correlation coding of shape structural information.Furthermore,it can avoid the influence of local differences between hand-rendering sketch and view edge.Finally,the algorithm encodes the features with hash regularization constraint and cross entropy loss function to avoid the divergence of eigenvalues.The experimental analysis on SHREC13 and SHREC14 shows that the retrieval accuracy of hash self attention end-to-end network is better than the existing typical algorithms.The mean average precision performance is improved by 6%.