摘要
深度学习技术在立体匹配算法领域中表现出良好的性能,现有方法通常使用类U-Net结构进行特征提取,并且仅利用3D卷积进行成本聚合.然而,类U-Net结构的下采样过程可能会丢失图像的某些细节信息,而3D卷积的操作忽略了图像整体结构信息,使算法在无纹理和重复纹理区域效果不佳.此外,现有模型往往难以同时保证算法的精度和实时性.针对以上问题,本文构建了一个基于3D平面向注意力的实时立体匹配网络,该网络首先利用上下文细节融合模块对类U-Net结构生成的特征进行补充,以增强特征表达能力.其次,通过基于3D平面向注意力的成本聚合,在有效获取全局信息的同时极大地降低了模型的计算量.最后,利用视差细化模块解决了因感受野过大导致的视差图边界平滑问题.实验结果表明,所提算法在保证实时性的同时,较大幅度地提高了模型的匹配精度.
Deep learning techniques have shown good performance in the field of stereo matching algorithms,but existing methods usually use U-Net-like structures for feature extraction and 3D convolution only for cost aggregation.However,the down-sampling process of the U-Net-like structure may lose some detail information of the image,while the 3D convolution operation ignores the overall structural information of the image,making the algorithms ineffective in untextured and repetitive texture regions.In addition,it is often difficult for existing models to guarantee the accuracy and real-time performance of the algorithm at the same time.To address the above problems,this paper constructs a real-time stereo matching network based on 3D plane direction attention,which firstly complements the features generated by the U-Net-like structure by using the context and detail fusion module to enhance the feature expression ability.Secondly,through cost aggregation based on 3D planar attention,the computational complexity of the model is greatly reduced while effectively obtaining global information.Finally,the disparity refinement module is used to solve the problem of disparity map boundary smoothing due to the large sensory field.Experimental results show that the proposed algorithm improves the matching accuracy of the model more substantially while ensuring real-time performance.
作者
吴期荃
魏国亮
刘舒婷
WU Qiquan;WEI Guoliang;LIU Shuting(College of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Business School,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《小型微型计算机系统》
北大核心
2025年第11期2683-2691,共9页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(62273239)资助
上海市“科技创新行动计划”国内科技合作项目(20015801100)资助。
关键词
立体匹配
深度学习
注意力机制
信息融合
stereo matching
deep learning
attention mechanisms
information fusion
