Enhancing palmprint image quality using convolutional neural network-based super-resolution methods

This thesis investigates the performance of deep learning-based super-resolution methods to enhance the quality of biometric images used in palmprint recognition systems. Widely used architectures in the literature—such as GAN, DCGAN, BGAN, BEGAN, and particularly the Adversarial Autoencoder (AAE)—were evaluated. The resulting super-resolved images were analyzed using both reference-based metrics like PSNR and no-reference metrics such as BRISQUE. Among these models, the AAE architecture achieved the best results; thus, the development process focused on this structure. The decoder component of the model was redesigned with convolutional layers and optimized to achieve 2× resolution enhancement. The model was trained on the IIT Delhi palmprint dataset. Experimental results demonstrated that the image quality enhancement process not only improved visual quality but also significantly increased biometric recognition accuracy.