@article{ZHANG2023, 
author = {Yan ZHANG and Changkang XU and Liqing CAO and Nian WANG and Mengyan WU and Yi XIA},
title = {Cross-Domain Pressure Footprint Images Retrieval Based on Mutual Information Disentangled Representations},
year = {2023},
journal = {Journal of South China University of Technology (Natural Science Edition)},
volume = {51},
number = {5},
pages = {78-85},
keywords = {image retrieval, cross-domain pressure footprint, disentangled representation, domain-shared representation, mutual information},
url = {https://www.sciopen.com/article/10.12141/j.issn.1000-565X.220572},
doi = {10.12141/j.issn.1000-565X.220572},
abstract = {As one of human biometric features, footprint is of great significance in the field of biometric identification. However, the pressure footprint images of different shoe types for the same person have significant differences in the footprint contour features, leading to large intra-class differences. For cross-domain retrieval of pressure footprint images, this paper proposed a cross-domain pressure footprint images retrieval method based on mutual information disentangled representations. Firstly, a multi-domain pressure footprint dataset containing 200 people’ s footprint images was constructed and the characteristics of cross-domain pressure footprint images were analyzed from qualitative and quantitative perspectives. Secondly, two independent encoders were used to construct an image disentanglement module, which disentangles the pressure footprint images into a domain-specific representation and a domain-shared representation, and ensures that the domain-specific representation contains more domain-related information through domain classification. Then, the distance between the domain-specific representation and the domain-shared representation was enlarged by minimizing mutual information loss. At the same time, in order to avoid the loss of information in the disentangled process, the original pressure footprint image was reconstructed based on the domain-specific representation and the domain-shared representation. Finally, the deep convolution features of the domain-shared representation were further extracted by feature extraction module and the cross-domain pressure footprint images retrieval was realized through the metric module which calculates the correlation degree between different features. The results of comparison and ablation experiments show that the disentanglement module of this method is effective and performs well on multi-domain pressure footprint dataset. The retrieval accuracy of the first query result reached 79.83%, and the average accuracy reached 65.48%.}
}