Automated identification of impact spatters and fly spots with a residual neural network

In criminal investigations, distinguishing between impact spatters and fly spots presents a challenge due to their morphological similarities. Traditional methods of bloodstain pattern analysis (BPA) rely significantly on the expertise of professional examiners, which can result in limitations including low identification efficiency, high misjudgment rates, and susceptibility to external disturbances. To enhance the accuracy and scientific rigor of identifying impact spatters and fly spots, this study employed artificial intelligence techniques in image recognition and transfer learning. Two types of bloodstains obtained from simulation experiments were utilized as datasets, and a pre-trained neural network, ResNet-18, was employed for feature extraction. The original fully connected layer was replaced, and a new fully connected layer with a dimensionality of 2 was introduced to fulfil the task requirements. The results demonstrate that the transfer learning network model, based on ResNet-18, achieved a maximum accuracy of 93 % in morphologically identifying impact spatters and fly spots. The objective is to assist crime scene investigators and BPA analysts to identify bloodstains at homicide scenes conveniently, rapidly and accurately, thereby furnishing scientific evidence for scene reconstruction and advancing BPA toward intelligent practices.