Shoeprint retrieval network based on composite attention mechanism

Shoeprint retrieval is an important component of footprint identification, where it searches for shoe information with the same footprint texture in a shoe database based on the footprints at the crime scene. In most cases, criminals do not leave their shoes at the crime scene, which means that a shoeprint retrieval is necessary to identify the suspect based on their footprints. However, due to the complexity of crime scenes, shoeprint images often suffer from problems such as blending of the shoe sole patterns with the background, image fragmentation, and acquisition differences, making accurate matching challenging. To address these issues, this paper proposes a shoeprint retrieval network based on a composite attention mechanism (CASN) that features a decoupled and disentangled enhanced position attention module (DEPAM) and a grouped channel attention module (GCAM). Specifically, the DEPAM module decouples highly coupled input information into salient texture information and whitened area similarity information, while the GCAM module captures long-range contextual information in the channel dimension. We introduce a spectral difference orthogonality regularization term to reduce the correlation between features, and then describe the network training method and parameter settings. Finally, by conducting comparative and ablation studies, we confirm the effectiveness of our proposed algorithm and its components, with our CASN achieving outstanding performance on the FID-300 dataset, securing top cumulative matching scores across the board except for Rank1%.