基于联合注意力-可分离卷积的立体匹配算法

Stereo Matching Algorithm Based on Joint Attention-separable Convolution

为解决立体匹配网络模型轻量化与高精度不能共存的问题,提出新的立体匹配算法CSA-Net(attention and separable convolution network)。该算法是在特征提取阶段,利用类ResNet进行特征提取,训练空洞金字塔池化(atrous spatial pyramid pooling,ASPP)模块扩大感受野,提取多尺度上下文信息,加入联合注意力机制(channel ang spatial attention module,CSM),在空间和通道维度提高表征能力,关注重要特征并抑制不必要的特征。在特征融合阶段,将2D深度可分离卷积提升到3D来代替原网络中标准3D卷积在空间维度和通道维度分别进行卷积运算,以降低特征融合网络的参数量与模型运行时间。实验结果表明,本文所提出的立体匹配网络模型在KITTI 2012和2015数据集进行验证,在三像素匹配误差率为1.44%和2.24%,模型运行时间减少近1/3。因此,相比于其他算法实现了更高的匹配精度和更快的运行速度。

A new stereo matching algorithm called CSA-Net was proposed to solve the problem of the coexistence of lightweight and high-precision in stereo matching network models.In the feature extraction stage,ResNet-like feature extraction was employed and the empty pyramid pooling(ASPP)module was trained to expand the receptive field and extract multi-scale context information.Additionally,a joint attention mechanism(CSM)was added to improve the representation ability in the spatial and channel dimensions by focusing on important features and suppressing unnecessary ones.In the feature fusion stage,2D depth separable convolution was extended to 3D to replace the standard 3D convolution in the original network to reduce the parameter amount and model running time of the feature fusion network.The experiment result validates the stereo matching network model proposed in this article on the KITTI 2012 and 2015 datasets,with the three pixel matching error rates being 1.44%and 2.24%,respectively.The model running time is reduced by nearly 1/3,leading to higher matching accuracy and faster running speed compared to other implementations.