Differences in shooting devices,angles,and lighting,as well as the interference from the similar targets,severe challenges to the task of cross-device object identification are posed.Aiming at the problems of intra-class differences and inter-class similarities in the identification process,an identifica-tion model that combines multi-scale and frequency domain features is proposed.An attention mechanism to the backbone network is added to improve the attention of the model to high-discrimination features.In the branch network,an attention-based multi-scale expansion and fusion module is designed to perform multi-granularity sampling fusion on different depth features to enhance the spatial mapping ability of the network.A self-learning frequency domain convolution module is constructed to realize the fusion of multi-scale and frequency domain features during the post-processing phase,and the frequency domain information is used to improve the accuracy of measuring the similar targets.After experiments,the aver-age mean accuracy(mAP)and first hit precision(Rank-1)of the algorithm in the Veri776 and Vehicle ID datasets are 81.60％,97.20％,90.50％,and 85.30％,respectively,and the results are better than those of the mainstream methods in recent years.And it can meet the requirements of multi-target identification of cross-equipment.
维普
万方数据