摘要
针对街景影像中的重复纹理和弱纹理区域,深度学习方法易受到内部特征相似性较高以及特征模糊等因素的影响,导致区域内部匹配精度较低问题,该文提出一种具有上下文感知的大支持窗立体匹配网络FELNet。针对重复纹理区域内部特征相似性较高的特点,引入混合注意力模块(CBAM),注意力机制使网络能够更加专注于区域内部的显著特征,从而有效地提升了区域内部特征的捕捉和表达,以此来增强重复纹理区域的特征区分度;针对弱纹理区域特征模糊的问题,提出了多尺度上下文增强模块(MCE)和H-D,两个模块在代价聚合中以串联形式处理,通过这种方式能够在保留上下文信息的同时,提升特征的质量和准确性,从而解决弱纹理区域精度下降问题。实验结果表明,相对于其他方法,该文提出的FELNet模型在匹配结果上取得了最佳结果。
When dealing with repetitive and weakly textured areas in street scenes,deep learning methods are susceptible to factors such as high internal feature similarity and fuzzy features,resulting in low matching accuracy within the area.This article proposes a context aware large support window stereo matching network FELNet.In response to the high similarity of internal features in repetitive texture regions,the convolutional block attention module(CBAM)is introduced,which uses an attention MCEhanism to enable the network to focus more on prominent features within the region,effectively improving the capture and expression of internal features in the region,thereby enhancing the feature differentiation of repetitive texture regions;To address the issue of feature blurring in weakly textured areas,the multi scale context enrichment module(MCE)and dilated hourglass module(H-D)are proposed.The two modules are processed in series in cost aggregation,which can improve the quality and accuracy of features while preserving contextual information,thereby solving the problem of decreased accuracy in weakly textured areas.The experimental results show that the FELNet model proposed in this paper achieves the best matching results compared to other methods.
作者
戴激光
罗方泽
DAI Jiguang;LUO Fangze(School of Surveying,Mapping and Geographic Sciences,Liaoning Technical University,Fuxin,Liaoning 123000,China)
出处
《测绘科学》
CSCD
北大核心
2023年第12期234-242,共9页
Science of Surveying and Mapping
基金
国家科学自然基金项目(42071428)
关键词
深度学习
重复纹理
弱纹理
注意力机制
上下文信息
deep learning
repetitive texture
weak texture
attention MCEhanism
context information
