Hybrid Super-Resolution and Keypoint For Forensic Tattoo Recognition

This study presents a hybrid processing pipeline that combines super-resolution techniques with keypoint-based feature extraction for forensic tattoo imagery, focusing on regions of interest whose smallest spatial dimension is below 100 pixels. The pipeline integrates three super-resolution models (EDSR, ESRGAN, and SRCNN) with three classical keypoint detectors (SIFT, SURF, and AKAZE). The evaluation is conducted on the BIVTatt dataset, collected in a forensic context in collaboration with a national investigative police agency in Costa Rica. The analysis incorporates image-quality metrics (PSNR and SSIM), keypoint based indicators (detected keypoint counts, relative variation with respect to downscaled inputs, and recovery ratios with respect to original resolution images), and detection-related metrics computed using the original-resolution images as reference (precision, recall, and F1-score), all derived from count-based formulations. Results are reported for ten representative low-resolution tattoo samples at the individual level, together with aggregated statistics across all evaluated cases. The super-resolution methods exhibit distinct numerical behaviors in terms of reconstruction metrics and keypoint detectability, while the keypoint detectors show heterogeneous responses to super- resolved inputs and low-contrast regions. All measurements are presented descriptively, without inferential claims. The processing workflow and evaluation protocol are fully documented to support reproducibility and to provide a quantitative reference for future studies on forensic tattoo imagery under reduced-resolution conditions.