基于注意力机制的立体匹配算法  

STEREO MATCHING ALGORITHM BASED ON ATTENTION MECHANISM

作　　者：黄怡洁 朱江平[1] 杨善敏 Huang Yijie;Zhu Jiangping;Yang Shanmin(Sichuan University,Chengdu 610065,Sichuan,China)

机构地区：[1]四川大学,四川成都610065

出　　处：《计算机应用与软件》2022年第7期235-240,309,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61901287);四川省重大科技专项(2019ZDZX0039)。

摘　　要：为了提升遮挡区域等视差估计精度,提出基于注意力机制网络的立体匹配算法(AMSN)。在特征提取网络部分,给残差网络中浅层分配更多的块数以获取细节信息,并用注意力模块分别从空间和通道维度捕获全局信息,使得提取的特征包含丰富的语境信息。在代价聚合模块采用3D卷积并利用编码解码结构进行多尺度的特征融合。其在SceneFlow数据集上的EPE和在KITTI数据集上的3像素误差指标均优于现有性能较好方法,实验结果证明了该方法在立体匹配上的有效性。In order to improve the accuracy of disparity estimation in areas such as occlusion, a stereo matching algorithm(AMSN) based on attention mechanism network is proposed. In the feature extraction network part, shallow layers in the residual network were allocated more blocks to obtain detailed information, and adopting an attention mechanism module to extract global features from channel and spatial dimensions respectively, so that the extracted features contained rich contextual information. The cost aggregation module used 3 D convolution and used the encoding and decoding structure to perform multi-scale feature fusion. The EPE on the SceneFlow dataset and the 3-pixel-error on the KITTI dataset were better than the existing methods with good performance. The experimental results prove the effectiveness of the proposed method in stereo matching.

关 键 词：立体匹配 注意力机制 深度学习 卷积神经网络 

分 类 号：TP390[自动化与计算机技术—计算机应用技术]