| Blood Pattern Analysis (BPA) is a core forensic method for reconstructing crime scenes, which traditionally relies on a trigonometric relationship to estimate a bloodstain’s angle of impact. Manual methods are labor-intensive and prone to human error, particularly with irregular boundaries. Prior work in this field, which attempted to automate BPA using baseline Artificial Intelligence, achieved approximately 78% accuracy, but identified a crucial performance gap in the 60°-90° range. In this range, bloodstains are nearly circular, making it difficult to distinguish the directional tail required for accurate Estimation. To address this challenge, this study proposes an automated deep learning framework utilizing a Convolutional Neural Network (ResNet34) augmented with a Convolutional Block Attention Module (CBAM) and MixUp augmentation. ResNet34 provides effective feature extraction while alleviating gradient vanishing issues. However, it treats all regions of an image uniformly. CBAM complements this architecture by dynamically reweighting spatial and channel-wise features, enabling the model to focus on subtle edge elongations and suppress irrelevant background noise. In addition, because blood droplet geometry changes continuously across adjacent impact angles, MixUp is incorporated to help the model learn smoother transitions between neighboring classes and reduce rigid categorization errors. Unlike earlier approaches that processed multiple stains together, our pipeline isolates individual droplets. This single-stain strategy increases the effective dataset size and more precisely captures the region of interest. Using standardized 10-degree bins (10°-90°), the proposed model achieved a strict classification accuracy of 56.99% and a forgiving (+/- 10°) accuracy of 86.06%. In the particularly difficult 60°–90° range, the model achieved an exceptional forgiving accuracy of 94.12%. These results show that the proposed framework improves interpretation of subtle high-angle bloodstain patterns and serves as a robust decision-support tool for forensic experts. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.