Study On Adaptive Image Denoising Algorithm Based On Image Quality Assessment

In imaging process, acquired images are inevitably contaminated by noise produced by thermal fading of light sensors or transmission process and other factors. Noise is one of main reasons that affects the image quality and visual perception, it seriously reduces the application value of acquired images and interferes subsequent high level image processing, like image restoration, visual tracking, image registration, image segmentation and image classification and so on. The study of image denoising has fundamental and far-reaching significance for improving image quality and meeting the needs of further image processing.This paper first employs the widely used Natural Scene Statistics(NSS)method to gain the image feature samples by extracting wavelet features from natural image set. Then, it takes feature samples as inputs for machine learning system to build noise type classification model and noise level prediction models. Finally, it enables different image denoising algorithms to have adaptive parameter setting mechanism by taking the advantage of the gained noise level prediction models.First, the noise type classifying accuracy of natural images contaminated by Gaussian Noise and Impulse Noise reaches a good level, by Mixed Noise, it comes to a level that outperforms manual classification. Second, the noise level predicting precision of natural images contaminated by Gaussian Noise and Impulse Noise is excellent. Third, when got the right noise type, it chooses the corresponding prediction model to enable the adaptive parameter setting mechanism of CBM3 D,NAFSM and NCSR. And last, the advanced Full-reference image quality assessment methods called SSIM and FSIM are used to evaluate the denoising performance, experimental result shows that adaptive CBM3 D and NAFSM denoising algorithms have overall improvement of noise suppressing, image details and feature preserving in every noise level, and for adaptive NCSR, despite affected by its own internal parameter regulatory mechanism, it still has overall improvement in low noise level. This prediction approach is actually a No-reference IQA, its superiority is that its assessment output is the noise estimation of the contaminated image and the estimation deviation is very low; it could be applied to a wide range of denoising algorithms that have fixed parameter setting.The experimental result shows that the classification model and prediction models have great classifying and predicting performance when tested by White Gaussian Noise and Salt & Pepper Noise, and successfully improved the performance of denoising algorithms.