Research On Non-local Means Image Denoising Algorithm Based On Features

The images corrupted by noise not only bring about poor visual effects but also lose part of the useful information, which is not conducive to the subsequent image processing and analysis. As a critical link of digital image pre-processing, image denoising is a popular but difficult problem in the field of digital image processing, which can improve the image quality.Non-local means(NLM) filter algorithm which takes full use of the self-similarity of images is currently a image processing method with strong ability of denoising. Because of the high computation, it is difficult for NLM algorithm to meet the needs of high processing speed in practical applications. Therefore, the main task of this thesis is to explore a fast NLM algorithm which can effectively keep the image denoising effect and realize its GPU parallel implementation.The main research contents are as follows:(1) Aiming at reduction of calculation quantity of the NLM, an NLM method based on adaptive search window(NLM-Adaptive Window, NLM-AW) is firstly proposed. A wide search range in the edge region and a smaller search range for non-edge region are used to reduce the computational complexity; then, in order to further improve the calculation speed, a fast step is adopted in horizontal and vertical direction(NLM-Fast Adaptive Window, NLM-FAW) for searching pixel patches. Experimental results show that the algorithm can effectively remove noise and the computation speed is improved by 17-66 times compared with that of the traditional NLM;(2) By combining with local binary pattern, a new local binary descriptor(LBD) is proposed. A binary description of the local structures of images can improve the robustness; and the logical operations take replacement of the Euclidean distance to measure the pixel block similarity, which can further increase the speed and protect the image edge at the same time.(3) An application of the proposed method in(2) is conducted on Optical Coherence Tomography(OCT) medical images. The comparation with other typical methods reflects the advantages in both denoising effects and computational time of the proposed method.(4) In order to realize the the algorithm for real-time process, a GPU parallel processing is also conducted. Experimental results on the Intel Core i5 processor and NVIDIA GeForce GTX660 show that, the computation speed is increased by 26 times. With a consistent effect with the current mainstream denoising method, the parameters of NLM-FLBD can be adjusted to achieve 51 frames per second, which can meets the requirements of real-time processing.