Retrieving shoeprint images using convolutional neural networks

Footwear marks are commonly found at crime scenes and can be used for investigation to identify potential suspects, and as evidence in court once a suspect has been charged with the crime. Shoeprint image retrieval refers to searching a database of shoeprint images taken from suspects for matches to a shoemark from a crime scene, to identify a potential suspect that could have left the shoemark. Recently, this task has been addressed by matching features computed using pre-trained convolutional neural networks. We build on this work by investigating further convolutional neural networks and usage of their various internal representations for shoeprint retrieval. We use three datasets to evaluate performance, and achieve state of the art results by matching feature maps extracted from EfficientNetV2 networks.