Cross-comparison of barefoot and sock-clad footprint evidence using an enhanced Siamese network approach

Traditional barefoot impression examination faces significant limitations in comparing impressions with sock-clad impressions. This paper presents the first cross-comparison study of barefoot and sock-clad impressions in challenging mixed datasets. We propose an enhanced Siamese network approach for the cross-comparison of barefoot and sock-clad impression evidence. Our methodology employs a dual-branch feature extraction framework based on ResNet34, enhanced with a channel-level generalized mean (GeM) pooling strategy and metric learning through hard sample mining. Research utilized 800 right footprint samples from 800 participants, augmented with 800 left footprint samples generated through mirror transformation, totaling 1600 samples for evaluation. Experimental results demonstrate that the proposed method achieves 63.4% Top-1 accuracy and 90.9% Top-10 accuracy in challenging mixed retrieval environments. The ResNet34 architecture with improved GeM pooling showed superior performance compared to alternative network architectures and pooling strategies. This research addresses critical challenges in the comparison of sock-clad impressions to barefoot impressions, particularly for cases where perpetrators wear socks to minimize distinctive impression evidence and sounds in burglary, homicide, and other crimes, providing a more objective, quantifiable automatic comparison method for barefoot and sock-clad impression identification with substantial practical value for criminal investigations.