Integrated edge thresholded deep network for image denoising

Abstract

The outcome of the denoising network suffers from over-smoothing effect, due to this, the texture content of the object will be lost. Lack of accurate texture properties in the image may lead to inefficient object segmentation and classification. This paper proposes an edge-preserving thresholding approach and applies it to the output of the denoised network. The thresholding approach relies on the distance and weight factors, which move the noisy components toward the mean of the subspace. This proposal is meant to treat over and under-smoothed components, where the smoothing decrement or increment is controlled by the threshold calculated with the average mean of the components in the respective subspace. The approach is compared with state-of-the-art methods in terms of image quality, and it is observed that this approach increases the quality proportionately. The result depicts that there is a significant improvement in PSNR of about 0.7~ 1 dB with the proposed integrated mechanism when compared against the conventional CNN-based image denoiser. Moreover, the edge details are better preserved with the proposed integrated mechanism.