【摘要】 针对可见光-红外行人重识别任务中利用行人细节特征减轻可见光图像和红外图像之间的模态差异而导致的过拟合问题，提出了一种基于改进卷积神经网络(CNN:Convolutional Neural Networks)的结合数据增强技术与细节特征提取方法的端到端网络。以生成多样嵌入特征的同时匹配行人细节特征减轻模态差异为目标，设计了一个高效的双分支注意力模块学习信息更加丰富的特征表达，并提出了一种具有数据增强作用的三元组损失函数缓解过拟合。通过在公共数据集SYSU-MM01和Reg DB上进行的大量实验表明，所提出的方法优于其他方法，有效减轻了注意力机制导致的过拟合问题，提高了行人重识别的准确性。更多还原

【Abstract】 To solve the overfitting problem caused by leveraging pedestrian detail features to reduce the modality gap between visible and infrared images in the visible-infrared person re-identification task,an end-to-end improved CNN( Convolutional Neural Networks) based on data augmentation techniques and detail feature extraction methods is proposed. With the goal of reducing the modality gap by generating diverse embedding features while simultaneously matching pedestrian details,an efficient dual-branch attention module is designed to learn more informative feature representations,and a triplet loss function with data augmentation is proposed to alleviate overfitting. Extensive experiments on the public datasets SYSU-MM01 and RegDB demonstrate that the proposed method outperforms other methods,effectively mitigating the overfitting problem caused by attention mechanisms and improving the accuracy of person re-identification.更多还原