Crime Scene Analysis Using Deep Learning

Abstract

Gathering evidence is one of the key activities during a criminal investigation. Image processing can assist in going over the scene of crime, or even recreating it. It isn’t always possible to expect the crime scene to be untampered with. However, with visual documentation of crime scenes comprising of various objects detected, it could be possible to come back to the images. This could assist in looking for something that has been discovered in further developments of the case. Additionally, with CCTV footage capturing tons of potentially crucial information, it is difficult for law enforcers themselves to go through the footage. Crime scene images are hard to find, and no known dataset is available to train the models directly on crime scene images. However, other popular datasets for image captioning do possess some crime scene images. The proposed approach is to train the models on a general vast dataset, MSCOCO. If the crime scene images are directly passed to the models, they will grab the main idea and give the sentence implying a person is dead. The idea is, however, to capture the more minute details ignoring the obvious conclusion that a person is dead. For crime scene images, the images are segmented into 9 images and passed individually to the model. This will generate 9 sentences for the input crime scene image with insight into the crime scene as opposed to captioning the crime scene. The three deep learning models proposed to use for generating sentences are: Inceptionv3-LSTM network, VGG-16-LSTM network and ResNet-50-LSTM network, and 0.1771, 0.11 and 0.1784 are the respective BLEU scores obtained. The users preferred the Inceptionv3-LSTM model over ResNet-50-LSTM model with a difference of 14.8% votes.