Document Type : Original/Review Paper
Authors
Department of Electrical Engineering, Imam Khomeini International University
Abstract
This study presents CSA-TopDB, a network aimed at improving the efficacy of person re-identification. Building on the Top-DB framework with a ResNet-50 backbone, CSA-TopDB incorporates a channel attention mechanism that facilitates the extraction of more distinctive and detailed features. The primary innovation is a cascaded channel attention mechanism, where global channel attention is first applied to the feature maps, followed by channel-wise splitting and independent local channel attention for each channel group, which systematically refines the features by transitioning from broad global context to fine-grained local details. To implement channel attention, two representative channel attention mechanisms, namely the Squeeze-and-Excitation (SE) and Efficient Channel Attention (ECA), were investigated. Based on the experimental results, ECA was adopted in the final CSA-TopDB model because it achieved a better balance between recognition performance and computational efficiency. The inclusion of global processing and regularization branches further improves the network's robustness and generalization. Experimental findings across Market1501, DukeMTMC, and CUHK03 (L) datasets indicate consistent improvements compared to the baseline method (Top-DB), specifically improvements of 4.3% in mAP and 2.8% in R1 accuracy on the DukeMTMC dataset have been achieved. These findings confirm the efficacy of the introduced channel splitting attention mechanism for person re-identification.
Keywords
Main Subjects
)