基于灾难性遗忘及组合叠加擦除的跨模态行人重识别预训练方法OACSCDCSTPCD

To meet the need of building a 24-hour full-time video surveillance system,cross-modal pedestrian recog-nition based on visible light and near-infrared is widely concerned by industry and academia.However,most of the current cross-modal pedestrian recognition tasks attempt to use pre-trained models on ImageNet to learn modal commonalities in ad-vance,but there are large modal differences between ImageNet and cross-modal pedestrian data,in the pre-training process,the color information is taken as one of the distinguishing features,which leads to the common features learned in the pre-training is not suitable for the information representation of the colorless infrared image.This paper proposes a self-super-vised cross-modal pedestrian recognition pre-training method based on catastrophic forgetting and combined superposition erasure.Firstly,the pre-training data are filtered by using the proposed catastrophic forgetting score,the aim is to reduce the domain gap between the pre-training data and the follow-up task data,and further reduce the training time of the model.Secondly,aiming at the key distinguishing feature extraction in traditional cross-modal identification,this paper designs a strong channel data enhancement strategy by erasing and combining the R,G and B channels at the channel level,multi-type samples with different colors are generated,which makes the model focus on texture information instead of color infor-mation.Finally,a self-supervised learning framework for cross-modal tasks is constructed based on the cross-modal data filtering index and channel enhancement strategy.The experimental results show that the ResNet50 network trained by the proposed pre-training method is superior to the current self-supervised pre-training methods when migrating to a large num-ber of cross-modal pedestrian recognition methods,on the basis of AGW,Rank1 and mAP were increased by 8.02％and 5.81％respectively.