Pre-activated simple attention network with SoftPool for optical footprint image retrieval
摘要
There is a high degree of similarity between the footprints of different individuals, and the formation of footprints is easily affected by walking conditions, which poses a great challenge to footprint image retrieval. Therefore, the key to optical footprint image retrieval lies in mining and learning discriminative details in footprint images. This paper takes optical footprint images as the research object, collects and constructs a dataset of 23,200 footprint images containing 2,700 individuals, and proposes a footprint retrieval network framework based on pre-activated Simple Attention Module (SimAM) to solve the footprint retrieval problem from the perspective of fine-grained image retrieval. Firstly, the feature extraction module uses ResNet-50 as the basic network to extract effective features from optical footprint images. Secondly, the feature set is constructed by the extracted features in the feature embedding module, and the pre-activated SimAM is proposed to focus on the 3-D information and overall contour distribution of the footprint image, thus learning more discriminative optical footprint features. Finally, the dual constraint loss function and SoftPool are used for optimization to perform the footprint retrieval task. Compared with other image retrieval methods, the experimental results show that our method achieves an average Rank1 and mAP of 95.72