基于特征学习的双路径红外-可见光行人重识别算法

Dual-path infrared-visible person re-identification algorithm based on feature learning

由于拍摄视角、行人姿态的变化以及不同的相机光谱造成的额外跨模态差异,RGB图像和红外图像之间存在着明显的差异,提取有效的模态共享特征是红外-可见光行人重识别中的难点。本研究提出一种双路径学习算法来识别特征,利用改进的BNNeck模块来提取RGB和红外图像的特征信息,改善算法的识别性能。该算法首先将注意力机制引入双路径特征学习网络,获取RGB图像在空间维度和通道维度上的特征信息,实现红外特征信息匹配;然后,将BNNeck模块引入至跨模态行人重识别算法,减少模态特征信息差异,加快算法收敛速度;最后,在异质中心损失函数和交叉熵损失函数的基础上,引入跨模态下行人身份损失函数,提高行人识别的准确性。SYSU-MM01和RegDB数据集的实验结果表明,相对于目前大多数已有算法,所提算法具有更好的泛化能力和鲁棒性,Rank-1/mAP分别达到59.39%/85.44%和57.81%/73.19%,比最新算法分别提高2.43%/2.86%和2.44%/1.19%。

Due to the changes of shooting angle,pedestrian posture and additional cross-modality differences caused by different camera spectra,there are obvious discrepancy between RGB images and infrared images.A difficulty in infrared-visible person re-identification is how to extract effective modal shared features.A novel dual-path algorithm was proposed to learn recognition features and an improved BNNeck module was utilized to extract the feature information of RGB and infrared images,which can improve the recognition performance of the whole algorithm.Firstly,the attention mechanism was introduced into the dual-path feature learning network to obtain the feature information of RGB images in spatial and channel dimensions,which achieved infrared feature information matching.Then,the BNNeck module was introduced into the cross-modality person reidentification algorithm to reduce the cross-modality differences and accelerate the convergence speed of the algorithm.Finally,on the basis of hetero-center loss function and cross entropy loss function,the pedestrian identity loss function under different modes was introduced to improve the accuracy of person re-identification.The experimental results of SYSU-MM01 and RegDB datasets show that compared with most of the existing algorithms,the proposed algorithm has better generalization ability and robustness,with the mAP/Rank-1 reaching 59.39%/85.44%and 57.81%/73.19%respectively,2.43%/2.86%and 2.44%/1.19%higher than the latest algorithm respectively.