Enhancing face recognition in native low-resolution images using deep learning and bicubic interpolati

This paper addresses the challenge of face recognition in Low-Resolution (LR) images, mainly when the resolution is below 48x48 pixels, which is common in surveillance systems. Current face recognition algorithms struggle to deliver satisfactory results with such low-resolution images. This study utilizes over 16,000 face images with an average resolution of 20x20 pixels to improve recognition, applying deep learning and bicubic interpolation to enhance image resolution. Unlike traditional Super-Resolution (SR) methods that operate in the LR space, our approach introduces a novel data constraint that evaluates errors in the High-Resolution (HR) image domain. By leveraging the finer details in HR images, the reconstructed HR images significantly improve visual quality and recognition accuracy. This unique data constraint seamlessly incorporates discriminative features into the optimization process. Experimental results demonstrate that our method outperforms existing visual quality and recognition performance approaches.