An extremely randomized trees method for weapons classification based on wound patterns of sharp metals using ultrasound images

Abstract

The use of digital images has increased and great strides have been made in the area of forensic science, pathology and other technological discoveries, there is a great need for computational and forensic techniques capable of detecting objects in images for the purpose of criminal investigation through classification, regression and quantifying objects to establish linkages, associations, and reconstructions. This paper presents extremely randomized tree algorithm for weapons classification using ultrasound images of the wounds caused by stabbing with sharp metals, these ultrasound images describe the internal structure of the wounds with patterns which helps us in determining weapon used for stabbing in most of the homicidal cases. The proposed method uses ultrasound images as input, enhanced with preprocessing techniques and segmentation for the region of interest, the segmented image is used for extracting features such as texture, shape, and size of the wounds that will be data for extremely randomized trees algorithm. The methodology is trained and tested on the available database of 300 images, through a number of testing; the efficiency of the proposed methodology is tested rigorously. The testing results show that aside from exceptional cases, the classification methods are able to correctly determine with an average accuracy of 95.95%. Not many researchers worked on the said problem statement with the computational approach but the proposed method as compared to a traditional method in which recognition is by human verification manually, it is found that the proposed work has given higher accuracy results with the set of features selected for identification and classification.