A Comparative Study on Deep Learning Techniques for Tattoo Localization

Tattoo localization is a crucial task in forensic analysis, biometric identification, and visual recognition systems. While deep learning has significantly advanced object detection and segmentation, its effectiveness in tattoo localization remains underexplored. This study presents a comparative analysis of state-of-the-art deep learning models for tattoo localization, evaluating their trade-offs between accuracy, efficiency, and real-time applicability. A dataset of 5000 freely licensed tattoo images was curated, with 4000 images used for training and 1000 for testing. The models were trained and evaluated using deep learning frameworks, leveraging GPU acceleration to ensure optimal performance. The evaluation was conducted using key performance metrics, including mean Average Precision (mAP), Intersection over Union (IoU), precision, recall, F1-score, and processing speed measured in Frames Per Second (FPS), providing a comprehensive assessment of detection accuracy and efficiency. Results indicate that Faster R-CNN achieved the highest mAP@50 (0.5520) and IoU (0.7240), offering a balance between precision (0.9483) and recall (0.7333). Mask R-CNN demonstrated superior recall (0.96) but at the cost of increased false positives and lower processing speed (4.43 FPS). YOLO, despite being the fastest model (18.17 FPS), exhibited lower recall (0.37), affecting overall detection coverage. These findings highlight the strengths and limitations of current deep learning models for tattoo localization, offering valuable insights for forensic and biometric applications. Future work will explore alternative architectures, enhanced optimization strategies, and different computational frameworks to improve detection accuracy, efficiency, and real-time performance.