基于损失自注意力机制的立体匹配算法研究

Research on stereo matching algorithm based on loss of self-attention mechanism

针对现有的立体匹配算法在阴影、物体边缘和光照反射等区域匹配困难且存在大量错误结果的问题,设计了一种可拆卸的损失自注意力网络(loss self-attention net,LSAnet)查找图像中的匹配困难区域。LSAnet的网络各层相互稠密连接,应用了空洞卷积来扩大感受野,并以立体匹配算法生成的损失分布为标签,能够动态地进行有监督训练,最终生成匹配困难区域掩膜辅助立体匹配网络进行更好的优化;同时,改进了立体匹配网络中经典的特征匹配代价卷结构,降低了后续3D卷积的计算负荷,提高了匹配效率。实验结果表明,该算法相比于基准算法精度更高,并且可以提高算法对于匹配困难区域的鲁棒性。

Aiming at the problem that existing stereo matching algorithms is difficult to match in shadow,object edge and illumination reflection regions and has a large number of error results,this paper designed a detachable loss self-attention net(LSAnet)to search for the difficult matching regions in images.Each network layer of LSAnet densely connected with each other,it used atrous convolution to increase the receptive field,and carried out dynamic supervised training with the loss distribution generated by stereo matching algorithm as the label,and finally generated a mask for difficult matching areas to assist stereo matching network for better optimization.At the same time,it improved the classical feature matching cost volume structure in stereo matching network,which reduced the computational load of subsequent 3D convolution and improved the matching efficiency.Experimental results show that the proposed algorithm has higher accuracy than the benchmark algorithm,and can improve the robustness of the algorithm for matching difficult regions.