Study On Signal Dependent Noise Removal Based On BM3D Algorithm

Image is an important carrier of visual information, and digital image has become the main form of visual information with the rapid development of computer technology. In the process of its formation, storage and dissemination, digital images are subject to different levels of noise interference, affecting the acquisition and interpretation of the information. How to remove the noise from the images and how to get the information we need become the current hot areas of visual research. More importantly, in medical applications, because of the imaging method and the imperfection of imaging equipment, medical images contain complex noise, interfering with the diagnosis and treatment of diseases. Therefore, image denoising is extremely important with great theoretical and practical values.Based on whether the noise is related to the image or not, this paper introduces signal independent noise and signal dependent noise by comparing their differences. At the same time, a noise model is proposed, which can effectively represent different types of signal dependent noise. Then, the basic principles of digital image denoising and related algorithms are introduced. Finally, two medical imaging technologies are introduced, which are the X-ray phase contrast imaging and the optical coherence tomography (OCT) imaging degenerated by signal dependent noise, where signal dependent noise removal is pointed out as an important research topic in the field.Then, the relevant techniques and algorithms used in this article are introduced: the blockwise DCT algorithm and the empirical Wiener filtering algorithm. Meanwhile, a parameter estimation method of signal dependent noise is introduced, as well as several key techniques:the noise variance estimation by PCA, the selection of weak texture area, the parameter estimation by the maximum likelihood method. Our algorithm for signal dependent noise removal is based on these algorithms.The block-matching 3D collaborative filtering (BM3D) is one of excellent additive Gaussian noise removing algorithms. However, for clinical signal dependent noise, this method cannot be implemented for effective results. Employing the robust estimation of noise parameters and the accurate measurement of similarity, as well as the enhancement of noise removal with the self-similarity, we propose a series of signal dependent noise removal algorithms based on the BM3D algorithm. The innovations of this paper are mainly concluded in the following aspects:· Improved BM3D algorithm based on noise analysis. Signal dependent noise is characterized by its complexity and diversity in medical applications. Through estimating signal dependent noise parameters based on PCA and the maximum likelihood estimation, by which we adjust the threshold of DCT and the shrinkage of the Wiener filtering coefficient, the improved BM3D algorithm is proposed adaptive to different noise levels.· Similarity measurement based on structural similarity index. For signal dependent noise, measuring similarity by the classic Euclidean distance is not accurate enough. We measure the similarity of image blocks by the structure similarity index combined with the Euclidean distance, and then apply it to the collaborative filtering of the BM3D algorithm, enhancing the matching of similar blocks.· Enhancement of denoising processing based on self-similarity. For images degraded with heavy noise level, due to unreasonable coefficient shrinkage of the proposed algorithm, fake serious artifacts are obvious in images. Using the self-similarity smoothing for further enhancement of the denoising processing, we obtain better results of noise removal.We use the improved BM3D algorithms and the self-similarity enhancement algorithm for signal dependent noise to denoise the simulated and the real noisy images. In experimental results it is shown that the proposed method surpasses the classical BM3D algorithm and related denoising algorithms in both subjective visual effect and objective peak signal to noise ratio to a certain degree. In the denoising of the X-ray phase contrast image and the optical coherence tomographic image, the proposed algorithm has some advantages in maintaining image details and features.Advanced techniques such as the block matching collaborative filtering based on noise analysis and the structural similarity index, and self-similarity enhancement algorithm deepen and enrich the denoising processing of signal dependent noise, which can be applied for medical imaging and have great value in both theoretical research and clinical applications.