Identification of Crime Scene Shoe Imprints using Deep Learning Techniques

Abstract

Shoe prints are most commonly found evidence at crime scenes and can be used in identifying the suspects and in further investigation. A shoe print can provide different types of information which can be used for identification purpose i.e. shoe size, shoe brand, association of suspect to crime scene and association between multiple crime scenes. Identifying the crime scene shoe print is a challenging task due to the variability of found prints (including noisy and partial prints). Some existing approaches cannot perform well on large dataset and highly degraded prints. It is found that pre-trained convolutional neural networks are very effective in feature extraction and identification for this kind of domain i.e. finger print and palm print identification. This study proposes a method which uses convolutional neural network to improve the identification problem of crime scene shoe prints. This method extracts key features from shoe prints which are invariant to scale, rotation, occlusion and translation. In this proposed study we have employed pre-trained networks namely ResNet50, InceptionV3 and VGG16 on FID-300 data set for the identification of shoe impressions found from crime scene. The study reflected that ResNet50 achieved the highest accuracy of 96.31 %.