Improved Two-Branch Person Re-Identification Algorithm Based on Semantic Parsing and Saliency Detection

Person re-identification faces challenges from occlu sion, blurring, pose variation, and scale misalignment. We propos e SPSD-Net, an improved two-branch algorithm combining sema ntic parsing and saliency detection. The semantic parsing branch uses an enhanced ResNet-50 with IBN-Net and a channel attentio n module to segment body parts, then applies a Contextual Seman tic Information Fusion Network (CSI-FN) to integrate local and $\mathbf{g}$ lobal features at the pixel level. The saliency detection branch em ploys a CNN to capture the most salient pedestrian features and s upplements high-level semantics via CSI-FN to recover lost conte xtual information. By fusing complementary features from both b ranches, SPSD-Net mitigates occlusion effects. Extensive experim ents on Market1501 and DukeMTMC-reID demonstrate that SPS D-Net achieves superior recognition accuracy compared to existing models.