Comparative Evaluation of VDSR and Image Enhancement Techniques for Enhancing Low-Quality Latent Fingerprint Images

The quality of latent fingerprint images can vary significantly based on deposition conditions, enhancement techniques, and imaging environments. This study investigates deep learning-based enhancement techniques for improving low-quality latent fingerprint images. GAN-based methods were excluded due to concerns regarding potential ridge distortion and artifact generation, leading to the selection of the Very Deep Super-Resolution (VDSR) model, which is based on a CNN architecture specialized for image processing. Additionally, VDSR offers the advantage of processing fluorescent images without the need for preprocessing. The VDSR model was modified to enable direct input of low-quality images, and three versions were evaluated: the pre-trained model, a fine-tuned model trained with an additional 100 ink fingerprint samples, and a newly-trained model using a dataset of 5,000 ink fingerprints. The pre-trained model, which yielded the best results, was compared with other enhancement methods, including R-ESRGAN 4x, a GAN-based image upscaler from Stable Diffusion, and traditional techniques such as intensity index adjustment and spatial filtering implemented through ImagePro. In general, deep learning-based techniques outperformed traditional methods, with VDSR proving particularly effective for enhancing fingerprint images. A combined approach that integrates deep learning-based techniques with traditional methods is expected to hold promising potential for further improving the quality of latent fingerprint images.