Exploration of explicit regularization terms in deep image prior

Deep Image Prior (DIP) has emerged as an effective technique for various image restoration tasks, leveraging the implicit regularization properties of convolutional neural networks (CNNs). Despite its success, the role of explicit regularization in enhancing DIP remains underexplored. This paper investigates the significance of explicit regularization terms in DIP, with a focus on their impact on the quality and robustness of image restoration. We conduct a comprehensive analysis of different regularization techniques, including total variation (TV) and regularization by denoising (RED), to evaluate their influence on image denoising. Our experimental results demonstrate that explicit regularization significantly enhances DIP performance, offering improved consistency in denoising tasks. Furthermore, we examine the trade-offs between different regularization techniques regarding computational complexity and restoration quality. This study underscores the value of explicit regularization in enhancing DIP's effectiveness and provides practical guidelines for selecting suitable regularization strategies in image restoration applications. Notably, the experimental results indicate that incorporating explicit regularization variants of DIP, such as DIP-TV, can eliminate overfitting, resulting in a 2.7% increase in PSNR and a 1.8% increase in SSIM compared to the original DIP.