General image denoising framework based on compressive sensing theory

• Our proposed general image denoising framework improves many existing algorithms.
• Compressive Sensing theory can guide us to devise algorithms from our framework.
• For different cases, we devise four novel algorithms from our framework.

Image denoising is an important issue in many real applications. Image denoising can be considered to be recovering a signal from inaccurately and/or partially measured samples, which is exactly what compressive sensing accomplishes. With this observation, we propose a general image denoising framework that is based on compressive sensing theory in this paper. Most wavelet-based and total variation based image denoising algorithms can be considered to be special cases of our framework. From the perspective of compressive sensing theory, these algorithms can be improved. To demonstrate such an improvement, we devise four novel algorithms that are specialized from our framework. The first algorithm, which is for the synthetic case, demonstrates the considerable potential of our framework. The second algorithm, which is an extension of wavelet thresholding and total variation regularization, has better performance on natural image denoising than these algorithms. The third algorithm is a more sophisticated algorithm for natural image with Gaussian white noise. The last algorithm addresses Poisson-corrupted images. Compared with several state-of-the-art algorithms, our intensive experiments show that our method has a good performance in PSNR (peak signal-to-noise ratio), fewer artifacts and high quality with respect to visual checking.

We propose a general image denoising framework based on compressive sensing theory. Many existing denoising algorithms can be considered as special cases of our framework. We devise four specialized algorithms to improve these algorithms. Some results are shown as follows: Though this is a synthesis example, it demonstrates the potential of our framework. For other experimental results including natural images with Gaussian noise and Poise noise, please refer to the PDF of Manuscript.Figure optionsDownload high-quality image (319 K)Download as PowerPoint slide