Anisotropic Diffusion-Based Noise Reduction for Image Quality Enhancement

Abstract

Image denoising is the process of eliminating noise in a noisy image in order to improve the image quality. It is a fundamental pre-processing step for various image-processing applications such as image segmentation, video coding, and medical imaging. In this process, it is difficult to eliminate noise without any loss of structure information, such as edges or textures. Therefore, conventional image denoising methods have focused on the development of structure-preserving image denoising techniques that attempt to preserve the structure information when smoothing is applied to suppress image noise. Anisotropic diffusion (AD) is the most widely used approach to structure-preserving image denoising because of its high denoising performance with simple operations. Recently, various AD-based approaches have been proposed to enhance the denoising performance of typical AD by revising the diffusivity-selection method. Although these methods show the enhanced performance of denoising compared with typical AD, they still have the following drawbacks. First, the region analysis methods used in the existing AD-based methods cannot effectively distinguish weak structure information from image noise. Second, the existing AD-based methods cannot extract the optimal diffusivity for local regions because they only adjust the diffusivity proportionally based on the results of the region analysis. Third, the existing AD-based methods require that a region analysis to be performed iteratively before each AD filtering operation, thereby imposing high computational cost. Lastly, most of the existing AD-based methods cannot utilize the inter-color correlation, which can improve the quality of denoising, because they separately apply the diffusion process to each single color plane. To alleviate the above drawbacks of the existing AD-based methods, I propose two AD-based noise reduction methods: (1) dictionary-based AD and (2) extended-dimensional AD. In dictionary-based AD, a multiscale region analysis is proposed to enhance the accuracy of the separation between the structure information and image noise. Second, a dictionary-based diffusivity determination is proposed to choose the optimal diffusion threshold regionally using the classification result of the multiscale region analysis. Finally, I propose a single-pass adaptive smoothing using a diffusion path-based kernel (DPK) designed by approximating the iterative AD operations. It enables the proposed method to avoid the use of an iterative region analysis. In addition, the application of the dictionary-based diffusivity determination demonstrates greater effectiveness in enhancing the performance of noise reduction over the typical AD; this is achieved by utilizing the enhanced maximum smoothing strength of DPK compared with the typical AD. In the extended-dimensional AD, the dimension of diffusion of the typical AD is increased by utilizing inter-color planes. In this method, inter-color planes from different color planes are predicted by adjusting their local mean values such that they properly utilize an inter-color correlation for the AD-based noise reduction. Then, DPKs for the current color plane and the predicted inter-color planes (PIPs) are generated to transform the iterative AD into a single-pass smoothing, which can avoid the use of the iterative region analysis. Simultaneously, a regionally and directionally varying diffusion threshold is adopted for the current color plane to preserve image details and to improve the quality of noise elimination near strong edges. For the PIPs, diffusion thresholds are regionally adjusted depending on local correlations that exist between the current color plane and each of the PIPs. This is done to optimize the performance of noise reduction obtained from the extended diffusion dimension. Experimental results obtained using various test image sets show that the proposed dictionary-based AD improves the average peak signal-to-noise ratio (PSNR) and the average structural similarity (SSIM) by up to 0.90 dB and 0.092, respectively, compared with the existing AD-based methods. In addition to the quality improvements of denoised image, the dictionary-based AD effectively reduces the computational complexity by avoiding the use of an expensive iterative region analysis. In the proposed extended-dimensional AD, further improvements are achieved by increasing the average PSNR by up to 0.76 dB over that of the dictionary-based AD. Meanwhile, the computational complexity of the extended-dimensional AD is increased by only 12.4% compared to that of the dictionary-based AD.