A Synergistic Approach to Image Restoration DenseNet Enhanced Deep Image Prior

¹ Department of CSE, New Horizon College of Engineering, Bangalore, India
² Research Scholar-ICE, Anna University, Chennai, India
³ Department of IT, Karpagam College of Engineering, Coimbatore, India

^* Corresponding author: senthilanandhi.a@gmail.com

Abstract

The DenseNet-Enhanced Deep Image Prior (Dense-DIP) model employs a combination of theories from DenseNet architecture and Deep Image Prior framework to achieve the best results in terms of image restoration. Such a cutting-edge method utilizes the densely connected layers of DenseNet for efficient recycling of features around the boned edges thus ensuring effective extraction of hierarchical features without compromising the finer details in the structures. The subtraction of Dense-DIP from pre-trained networks and large datasets for labeling is possible as the method begins with a random noise network that is directly optimized onto distorted images. Important elements are also skip connections and early stopping, which often prevent overfitting and improve the quality of restoration. The model has broad applicability, including image denoising, inpainting, super-resolution, deblurring with a large improvement in quantitative and perceptual quality.