基于哈希自注意力端到端网络的三维模型草图检索

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

为了提高草图和三维模型视图嵌入特征的聚类性,提出一种结合自注意力和哈希正则化约束的特征提取算法.首先将三维模型渲染得到二维视图集,并通过边缘检测在草图和视图之间建立统一的特征描述空间;然后在共享权重网络中嵌入自注意力层,通过结构信息自相关性编码提高草图和视图的聚类性,避免局部差异性对结果的影响;最后对特征进行哈希编码,并嵌入哈希正则化约束和交叉熵损失函数,避免特征值发散.对基准数据集SHREC13和SHREC14的实验结果表明,该算法在哈希自注意力端到端网络的检索准确率方面优于已有的典型算法,平均准确率性能提高了6%.

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%.