A Judicious way to restore random impulse noise using iterative weighted total variation diffusion technique

Abstract

Various types of pixel candidates are available in the literature to replace impulse noise after effective detection. However, using them in the correct location and preserving the signal content, structural similarity, and image details is a task that draws attention, especially in a highly corrupted image. Non-linear Diffusion-based restoration is an efficient solution since it can iteratively update corrupted pixels without diffusing the edge. This work assigns the iterative weighted total variation diffusion technique only for the possibly noisy pixels in high noise ratio processing windows where the windows are pre-classified as low or high noise ratio by a custom CNN classifier. The work, called as CNN-based locally adapting filter (CNN-LAF), can achieve a high structural similarity of .9167 by maintaining a PSNR of 24.01 dB at a 0.8 noise ratio.