Research On Denoising Algorithm Of Images Based On Embedded System

Under the harsh condition or equipped with some bad quality of camera, device-perceived images usually go through problems like noises contamination, contrast degradation and so on, which will heavily affect the results of other operations of images processing, such as Face Recognition, Target Tracking, License Plate Number Recognition etc. So it's very meaningful to conduct a profound research on the image denoising technology. The non-local means algorithm takes full advantage of the pixel similarity and redundant information contained in the image, like the texture images or the structure similarity. By computing the similarity and evaluating pixels, the NLM algorithm can restore the noises contamination of images perfectly. The experiment turn out that The NLM algorithm outperforms other current denoising algorithms. But this method has some severe disadvantages, like high computational complexity, low veracity of similarity computation, and the fixed parameters to all the pixels etc. all those the inherent defects have produced a severe bad effect on its application and extension.In this thesis, we have a profound research on the Zernike-based NLM(ZNLM) filter, and try to improve its effect in the denoising quality. Then we analysis this algorithm and apply it to the embedded system based on FPGA and DSP. The achievements and works are summarized as below:(1) We proposed an Iterative Method Noise-Based NLM algorithm. The Method Noise is the difference of original image and the denoised image, which include a mass of efficient information, and we try to use the ZNLM filter to abstract the information from the Method Noise and put it back to the first denoised image. The experiment turn out that our idea is right. The PSNR increases about 4db in the first time iteration. Abstract efficient information from the Method Noise constantly and we get a better result. The proposed method outperforms the ZNLM algorithm in both visual quality and PSNR.(2) We accomplish and accelerate ZNLM denoising algorithm on the embedded system based on FPGA and DSP. By controlling the parallelized pipeline of FPGA module, DSP can optionally adjust the data quantity of exchange between FPGA and DSP, which make sure that the gigabit Ethernet bandwidth satisfies the task's requirement, and transfer the original image and its Zernike moment from FPGA to DSP correctly and efficiently. The parallelized pipeline speeds up the calculation of Zernike moments notably, and the restore image is transferred to PC through the JATG port on the embedded demoboard. We build an integral system including the image capture module, image processing part, and the image display part. Compared to MATLAB code running on PC, our embedded system achieves an acceleration rate of 340 or more.