结合分组相关性和注意力机制的立体匹配算法

Stereo Matching Algorithm Combining GroupingCorrelation and Attention Mechanism

目前立体匹配网络通常面临图像特征获取不充分、语义信息丢失和速度慢等问题。为了改善网络的特征获取能力和运行效率,提出一种结合多重注意力机制和分组相关性代价卷的立体匹配算法。首先,采用基于双重注意力机制的策略,通过融合可变形卷积构建多尺度特征提取网络。通道注意力和空间注意力机制能够在不同分辨率的图像特征提取过程中充分优化特征信息。同时,引入可变形卷积可以自适应采样物体的形状和尺寸,从而提高计算效率。接着,在代价卷的构造过程中,采用分组相关性的方法,同时结合连接特征和相似度特征生成代价卷。这不仅减少了参数和计算量,还保持了语义信息的完整性。最后,通过交叉融合不同尺度下的代价卷,得到最终的代价卷,并经过视差回归得到不同分辨率的最终视差图。实验结果表明,该算法在KITTI2015数据集上取得了显著的成果。全区域误差率仅为3.02%,计算时间为0.16 s,充分展示了该算法在保持低计算复杂度的同时获得了优越的匹配效果。

Existing stereo matching networks frequently suffer from inadequate image feature extraction,semantic information loss,and slow processing speeds.To ameliorate the network's feature acquisition capabilities and operational efficiency,a stereo matching algorithm that combines multiple attention mechanisms with grouping correlation cost volumes was presented.Firstly,a multi-scale feature extraction network was constructed by fusing deformable convolution based on a dual attention mechanism strategy.The channel attention and spatial attention mechanisms can fully optimize feature information in the process of image feature extraction with different resolutions.At the same time,the introduction of deformable convolution can adaptively sample the shape and size of the object,thus improving the computational efficiency.Secondly,in the construction process of the cost volume,the method of grouping correlation was used to generate the cost volume by combining the connection feature and the similarity feature.This not only reduces the number of parameters and computation,but also maintains the integrity of semantic information.Finally,the final cost volume was obtained by cross-fusing the cost volume under different scales,and the final parallax map with different resolutions was obtained by parallax regression.Experimental results show that the proposed algorithm achieves remarkable results on KITTI2015 dataset.The error rate of the whole region is only 3.02%,and the calculation time is 0.16 s,which fully shows that the algorithm can obtain superior matching effect while keeping the computational complexity low.