Lightweight Person Re-Identification Network Based on Random Color Dropout and Self-Attention

In the domain of Person Re-IDentification (ReID) tasks, images of pedestrians often exhibit noticeable color variations due to diverse environmental conditions. Notably, the accuracy of retrieval results improves when some color information is disregarded. To mitigate the adverse effects of color variations on recognition performance, this research introduces the Random Color Dropout (RCD) data augmentation strategy, employing the Local Aggregated Grayscale Transformation (LAGT) as its foundation. The proposed strategy enhances model’s robustness against color deviations by equitably adjusting transformation weights across the RGB image’s three channels. Additionally, to minimize overall model parameters, we employ the lightweight and efficient OSNet as the backbone. Recognizing the network’s limited attention to feature details and the interference of intricate scenes on recognition, we incorporate a complementary cascade-type Self-Attention Module (SAM). This module effectively consolidates spatial and channel information within the feature map, ameliorating information deficits in features and augmenting their discriminative properties. On benchmark datasets, Market1501 and DukeMTMC-reID, our method attains performance metrics of 95.5% and 89.2% in Rank-1, and 87.7% and 77.2% in mAP, respectively. These results underscore the superior performance of the proposed method compared to the prevailing mainstream algorithms.