【作者】 杨延西； 孙俏； 张少华；

【Author】 YANG Yanxi;SUN Qiao;ZHANG Shaohua;Xi’an University of Technology,School of Automation and Information Engineering;

【机构】 西安理工大学自动化与信息工程学院；

【摘要】 针对基于深度学习的立体匹配算法在病态区域上匹配效果不理想的情况,提出了一种基于混合注意力机制的视差计算方法。以DispNet网络为基准模型,通过引入残差结构和基于矩阵乘法的代价计算方法对网络结构进行调整,将并行的空间和通道挤压激励模块嵌入到网络中,增强匹配代价体中重要的空间与通道信息,提高网络的学习能力。实验结果表明,在SceneFlow数据集上的端点误差为0.66;在KITTI数据集的非遮挡区域上的平均错误率为2.08%,在全部区域上的平均错误率为2.13%。更多还原

【Abstract】 Aiming at the unsatisfactory matching effect of the stereo matching algorithm based on deep learning on morbid regions, a parallax calculation method based on the mixed attention mechanism is proposed.Taking the DispNet network as the baseline model, the adjustment and performance optimization of the network structure come true with the help of residual structure and cost calculation method based on matrix multiplication, and the network includes not only the parallel space module but also channel squeezing excitation module, which not only enhance the important space and channel information in the matching cost, but also improve the learning ability of the network. The experimental results show that the end point error on the SceneFlow data set is 0.66; the average error rate on the unshaded area of the KITTI data set is 2.08%, and the average error rate on all areas is 2.13%.更多还原