«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年06期

页码:

158-165

栏目:

人工智能

出版日期:

2025-06-10

文章信息/Info

Title:

An Unsupervised Stereo Matching Algorithm Based on Feature Perception and Guidance

文章编号:

1673-629X(2025)06-0158-08

作者:

魏东; 郑博闻*; 王思雨

沈阳工业大学 信息科学与工程学院,辽宁 沈阳 110870

Author(s):

WEI Dong; ZHENG Bo-wen*; WANG Si-yu

School of Information Science and Engineering,Shenyang University of Technology,Shenyang 110870,China

关键词:

立体匹配; 无监督; 特征感知; 特征引导; Dropout

Keywords:

stereo matching; unsupervised; feature perception; feature guidance; Dropout

分类号:

TP391.41

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0025

摘要:

针对立体匹配算法在处理物体边缘、视差不连续等细节时面临的挑战,以及有监督算法对数据标注的高度依赖性,提出了一种特征感知与引导的无监督立体匹配算法。 该算法在生成器的编码器部分嵌入特征感知模块。 该模块结合残差网络的稳健性,确保了特征提取的稳定性,还结合空洞金字塔卷积网络的广感受野特性,有效地扩大了特征捕捉的范围,此外,还辅以软池化技术,以增强特征的层次性和丰富性,使算法能够更好地应对图像中的细节变化。 为进一步提升特征的表征能力,引入了特征引导模块,通过结合通道注意力和空间注意力机制,动态调整不同通道和空间位置的权重来有效聚焦于关键特征区域。 此外,在判别器中加入 Dropout 层,以随机丢弃部分神经元连接的方式促使模型训练更加稳定,避免过拟合情况发生。 为了验证算法的有效性,实验采用了 KITTI 2015 数据集进行评估。 结果表明,与其他经典算法相比,该算法在细节及区域的效果、精度方面均有提升。

Abstract:

Addressing the challenges faced by stereo matching algorithms in handling details such as object edges and disparity discontinuities,as well as the high dependency of supervised algorithms on data annotation,an unsupervised stereo matching algorithm with feature perception and guidance is proposed. This algorithm embeds a feature perception module in the encoder part of the generator. This module combines the robustness of ResNet to ensure the stability of feature extraction,integrates the wide receptive field characteristic of atrous spatial pyramid pooling (ASPP) to effectively expand the range of feature capture,and is supplemented by soft pooling techniques to enhance the hierarchy and richness of features,enabling the algorithm to better cope with detailed variations in images. To further enhance the representation ability of features,a feature guidance module is introduced,which dynamically adjusts the weights of different channels and spatial positions by combining channel attention and spatial attention mechanisms to effectively focus on key feature areas. Additionally,a Dropout layer is added to the discriminator to randomly discard some neuronal connections,promoting more stable model training and avoiding overfitting. To verify the effectiveness of the proposed algorithm,experiments were conducted using the KITTI 2015 dataset for evaluation. The results demonstrate improvements in detail preservation,regional effects,and accuracy compared to other classic algorithms.

常用功能

更新日期/Last Update: 2025-06-10