基于双重三池化注意力机制的PSMNet算法  

PSMNet algorithm based on dual three-pooling attention mechanism

作　　者：刘腾飞 林冬云 兰维瑶 陈岳航 LIU Tengfei;LIN Dongyun;LAN Weiyao;CHEN Yuehang(School of Aerospace Engineering,Xiamen University,Xiamen 361102,China)

机构地区：[1]厦门大学航空航天学院,福建厦门361102

出　　处：《应用光学》2025年第2期327-335,共9页Journal of Applied Optics

基　　金：国家自然科学基金(62273285)。

摘　　要：为解决小型纹理类物体的视差计算及三维重建问题,提出了基于双重三池化注意力机制的PSMNet-ECSA算法。通过在残差网络主干中嵌入通道和空间注意力两个维度,每个维度以平均、最大、混合池化的方式进行特征维度融合,在一定程度上防止了过拟合现象,从而增强网络信息提取能力和泛化能力。在实验环境和数据集一致的条件下,经过SceneFlow、KITTI2015和真实场景实验分析,相对比原始PSMNet算法,本文算法在平均绝对误差、阈值误差等指标取得了10%的提升;将该算法应用于鲍鱼重建三维点云模型,长、宽、呼吸孔等距离测量平均相对误差在3%以内,能够以自动化的方式测量并记录小型海洋类生物的生长情况,具有良好的实际应用价值。To address the disparity calculation and 3D reconstruction problems for small textured objects,a PSMNet-ECSA algorithm based on a dual three-pooling attention mechanism was proposed.By embedding two dimensions of channel and spatial attention mechanisms into the backbone of the residual network,each dimension was fused using average pooling,maximum pooling,and mixed pooling techniques to merge feature dimensions,which prevented overfitting to some extent,thereby enhancing the network information extraction and generalization capabilities.Under the conditions of consistent experimental environments and datasets,an analysis was conducted through experiments on SceneFlow,KITTI2015,and real-world scenes.Compared to the original PSMNet algorithm,the proposed algorithm achieves a 10%improvement in metrics such as mean absolute errors and threshold errors.Applying this algorithm to reconstruct three-dimensional point cloud models of abalone,the average relative errors in distance measurements such as length,width,and breathing holes are within 3%,which can automatically measure and record the growth status of small marine organisms,and has promising practical application values.

关 键 词：视差图 立体匹配 堆叠卷积 注意力机制 三池化 

分 类 号：TN201[电子电信—物理电子学] TP391.4[自动化与计算机技术—计算机应用技术]