Design And Implementation Of Parallel Image Processing Algorithms Based On GPU

In recent years, with the continuous development and progress in imaging technology, the amount of data within a digital image has become larger and larger. When processing large-scale images, serial algorithms can hardly satisfy the real-time requirement due to their low processing speed. GPU (Graphic Process Unit) has many advantages such as many calculation units, and shows huge advantage in processing speed compared to CPU. This chapter mainly studies the GPU parallel computing, and makes GPU acceleration study for several typical time-consuming image processing algorithms. The main research work is described as follows:Based on the fact that the non-local means denoising algorithm shows a slow speed when denoising noisy images,this paper makes an analysis of the NL-means denoising algorithm’s steps and computational complexity and draws the following conclusion: the step of computing weights between different pixels is an iterative process, which undertakes a high computational complexity and belongs to regular computation.In this paper a parallel strategy is devised and the estimation of every pixel is computed synchronously in GPU. The experimental results show that the parallel algorithm significantly improves the image processing speed when maintaining the performance of the serial algorithm.The PPB(Probabilistic Patch-Based) denoising method shows a good performance when processing SAR images, but it has limitations in processing speed due to its own computational complexity. In this paper, we make a analysis on the serial PPB algorithm, parallelize some time-consuming steps dealing with the weights between pixels, compute the part of the algorithm which cannot be parallelized within CPU,and alternately execute the computation between CPU and GPU, at the meanwhile, a quantitative research on data transmission and computing time within GPU is made. The experimental results show that, when our proposed parallel PPB denoising algorithm is imposed on large-scale SAR images, it shows significant advantage in processing speed.Finally, focusing on the problem that it is difficult for traditional feature selection method to get good feature sets for image classification, we propose a novel parallel feature selection method for images based on immune clone algorithm,which is called PICFS(Parallel Immune Clone Feature Selection)algorithm. With an adaptive cross and mutation strategy embedded, in every immune clone operation this algorithm can always effectively keep the search direction toward the one with less dimensions when keeping the advantages of the predecessor, and the parallel procession significantly improves the algorithm’s speed. Experimental results show that,compared to traditional methods, the PICFS algorithm can achieve the better feature sets,which are fewer in dimensions when higher in accuracy.