Parameter Selection for Denoising Algorithms Using NR-IQA with CNN.

In order to yield satisfied image after denosing processing, the process of error tracing is nearly necessary for parameter selection. In practice, usually the choice of such parameters is time consuming and empirically dependant when a ground-truth reference is inavailable. Although some successful methods have been proposed in recent research, they still require certain parameters to be set a priori for parameter optimization. These methods tend to be strongly reliant on restrictive assumptions on the signal and noise of input image. In this paper, we propose a framework of parameter selection, which is based on subjective perceptive evaluation and implementated by a well-designed convolutional neural network. The proposed algorithm has the following advantages: (1) consistents with subjective perception, (2) does not require a reference image available, (3) with better capability of robustness and generalization, (4) trims the number of iteration of denosing in parameter selection. Experimental results show that our algorithm outperforms other methods in parameter selection. Our parameter trimming framework saves the computation of iterative image denoising up to 74%.