Reserch Of The Image Denosing Algorithm And Its Parallelization Based On Mathematics Morphology And Transformation Domain

Image denoising is a fundamental and significant technique in image processing. During the procedure of image digitalization, it is inevitable to bring in some noise, which will have side effect on people's understanding about the image and become the main factor of affecting the quality of the image. Using the technique of denosing, we can extract required signal from complex signal, control the interfering signal and make the image clear and precise. The clarity of the image mainly depends on the clarity of the curve of image, tiny line and tiny features, it also depends on whether the areas change between features is smooth enough. As for the traditional ways for image denosing, in the process of Image denoising, the edge of the will becomes vague so as to the information of these areas can not be keep anymore. In order to solve this problem, the researchers proposed a new kind of nonlinear method which is based on mathematics morphology. This method firstly analyze the character of the image, then matching signals with the pre-defined structured cell, using this method, we can extract the signal we need, and depress the noise, this method also has the feature about parallelism and speediness.This thesis is based on the theory of mathematics morphology, research denosing algorithm on binary image, gray scale image, color image. Based these traditional algorithms, according to some innovation and improvement, we further enhance the quality of the image.Designing and implementing the serial generalized composite types filter for binary image denosing algorithm. ??to the traditional morphological filter just use single cell structure, the result can not contain feature of different dimensions, and this method can not filter completely, so as to the detail of the image is vague. In this paper, we adopt the multi-scale and multi-dimension cell structure for the binary image denosing algorithm, not only the noise in the image can be controlled efficiently but also the geometrical characteristics of the image can be kept quite well. Meanwhile this thesis uses several important features of generalized composite types filter such as translation invariance, increase by degrees and idempotence to decrease the image noise. Based on the denosing algorithm of gray scale image, this thesis adopt a kind of Contourlet transformation which combine the denosing method of Cycle Spinning technique and HMT transformation of morphology. Contourlet transformation is a very kind of denoting method for two dimension image ,it has several good qualities such as multi-distinguishability, local property and directivity. Besides, the Cycle Spinning technique can depress the fake Gibbs phenomenon which is caused by lacking of translation invariance for Contourlet transformation. On the basis of combining these two method, this thesis use the HMT transformation in the theory of mathematics morphology to extract and remove the noise in high frequency which is produced by Contourlet transformation. Compared with traditional WT threshold value denosing method, this method smooth the noise, keep more image edge and detail texture, so as to get a better visual effects.As for color image denosing algorithm, this thesis adopt modified Median Filter algorithm based on RGB??or space and the scalar pattern filter algorithm of multi-structured cells. The traditional Median Filter algorithm uses a unified processing method for all data, it sorts every pixel and finds the median, it results huge computation and high time complexity. But the improved Median Filter adopts the algorithm of processing signal and noise separately, it can use the update data when sorting, so it overcomes the defeat in the traditional median, decreases the time complexity and obviously improves the filtering effect. Meanwhile, according to human eye have different sensitivity to the tricolor, we process the tricolor with different size structure cell, with the comparison of adopting same size structure cell, during the process of denosing, this method can keep more color information and do not bring in side effect to visual effect.Based on the color image denosing algorithm, this thesis deeply research the feasibility of the parallelism for improved Median Filter algorithm, and implement a parallelization algorithm for it. Using this parallelization algorithm, the computing efficiency of the improved Median Filter algorithm has been improved dramatically.According to analyzing the result of several simulation experiments and denosing performance indicators(MAE, MSE, NMSE, PSNR), this thesis draws a conclusion that the various improved algorithm we has proposed have better flexibility, the denosing ability has been enhanced notably, ultimately, the feasibility and the validity haven verified.