Enhancing Image Perception Quality: Exploring Loss Function Variants in SRCNN

Convolutional Neural Networks (CNN) have made significant strides in the field of image super-resolution, with SRCNN being a classic exemplar. SRCNN employs Mean Squared Error (MSE) as its loss function and supervises the network on a per-pixel basis. Nonetheless, research into the mechanisms of visual perception and objective evaluation metrics for super-resolution images remains inadequate. To align subjective and objective assessments, this study employs perceptual metrics to evaluate the effect of image super-resolution. Perceptual metrics mimic the early transformation processes of the human visual system, utilizing the NLPD (Normalized Laplacian Pyramid Distance) between images and the original scene radiance to assess the super-resolution effect. Building upon this, we transition SRCNN's loss function from MSE to two components, MAE (Mean Absolute Error) and NLPD. Using MAE as a measure of content accuracy and NLPD as a measure of perceptual similarity, we reconstruct high-fidelity super-resolution images by minimizing the NLPD between images and the original scene radiance. Experimental results indicate that, compared to SRCNN and BICUBIC, the super-resolution images we obtain not only restore content realistically but also align more closely with human visual perception.