Abstract
In forensic science, blood is a crucial piece of evidence for reconstructing crime scenes. Identifying and classifying blood can help confirm a suspect’s involvement, though various chemical processes are employed to identify bloodstains at the crime scene. However, such processes may deteriorate the obtained material and interfere with further DNA analysis. Hyperspectral Imaging (HSI) is a promising noncontact technique that can be utilized in forensic science examination at crime scenes for body fluid classification, including bloodstain detection and classification. Therefore, this work demonstrates the use of Hybrid Inception networks for HSI data analysis for bloodstain recognition and classification. For testing and validation, we make use of a Hyperspectral-based Bloodstain dataset that is openly accessible. A variety of detection scenarios with differing degrees of complexity are incorporated in this dataset. It allows evaluation of how well machine learning techniques work in various backgrounds, acquisition environments, blood ages, and situations where additional blood-like compounds are present. We conducted blood detection experiments using this dataset. We use the proposed Hybrid Inception network to compare the findings against a variety of widely accessible cutting-edge deep learning models, including 3D CNN, Hybrid CNN, and the Inception model. We carefully evaluate the results and discuss each examined architecture, taking into consideration the limited supply of training samples. Experiments show that the modified Inception network is an efficient and accurate classification model.
Keywords: Bloodstain Classification, Crime Scene Investigation, Deep Learning, Spectral Information, Hyperspectral Imaging, Inception Network.