Classification of Bloodstain Spatter Patterns with Convolutional Neural Networks

Bloodstain Pattern Analysis is a forensic discipline focused on the examination of bloodstains at crime scenes with a view to assisting with the reconstruction of events. Analysis of the physical characteristics of individual bloodstains and the overall pattern can assist with determining the mechanism that caused them, a task referred to as classification. Currently, the analysis of bloodstain patterns is still a manual and somewhat subjective task. With error rates of 13-23% for classification of spatter patterns, a more objective approach is recommended. The application of CNNs to the task of classification of three distinct spatter patterns, impact, cast-off and expirated is explored. A dataset of high-resolution laboratory-generated pattern images was used for training. To capture the entirety of the pattern that contains numerous small bloodstains spread over large areas, these images were composite images, ranging in size from 7 to 81 megapixels. This created a challenge for training the CNNs, with preprocessing steps required to create appropriate training data. Several pre-processing techniques, including dilation, resizing and cropping, were investigated and used to train two pre-trained CNN models, ResNet-50 and VGG-16. Dilated and resized images representing the full extent of the bloodstain patterns reached accuracies of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$&gt; 95\%$</tex>, but <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$&lt; 90\%$</tex> for partial patterns created through cropping.