Joint feature enhancement and semantic correlation mining for occluded visible-infrared person re-identification

The primary difficulty in the Visible-Infrared Person Re-Identification (VI-ReID) task is the substantial discrepancy in modality between daytime color images and nighttime infrared images. Prior VI-ReID methodologies have primarily concentrated on the analysis of high-level features, while neglecting the significant information contained within lowlevel features. Moreover, existing data sets do not take occlusion into account, while occlusion is prevalent in real world environments. To address these challenges, a Joint Feature Enhancement and Semantic Correlation Mining Network is introduced. Initially, an Infrared Enhancement Module (IEM) and a Visible Enhancement Module (VEM) are incorporated to bolster the feature representation of infrared images through the application of Contrast Limited Adaptive Histogram Equalization (CLAHE), and to enrich visible images using various channel transformations. Subsequently, a Triple Semantic Correlation Mining Module (TSAM) is devised to collaboratively explore semantic correlations across different stages, utilizing spatial, channel, and multi-head self-attention mechanisms. Ultimately, Cross-modality Contrastive and Dual Center Loss is devised to make full use of the abundant semantic information in the features. Experimental results validate that the method proposed in this paper exhibits superior performance on two public datasets and outperforms existing VI-ReID methodologies on a self-constructed dataset.