Towards The Parallelization For Typical Remote Sensing Image Enhancement Algorithms Based On MIC Architecture

Image denoising is an important step for Remote Sensing(RS) image preprocessing in related applications, and has become a research branch in the field of RS image enhancement. Currently, many scholars devoted themselves to the study of image denoising, and proposed many image denoising methods. However, most of those methods are not very satisfactory since a lot of texture details are often lost in the denoising processing. NLM(Non-local Means) algorithm, proposed by Buades et al in 2005, can prevent loss of details to a certain degree. NLM algorithm is based on the selfsimilarity principle. As a result, it can make full use of the redundant information that universally exists in the image. During the searching process, the weights to measure the similarity between pixels are calculated based on the differences between the gray value vectors of the similar pixel slices.Currently, the NLM algorithm is one of the best image denoising algorithms for its capability to better keep the image’s texture details. However, the time complexity of the algorithm is very high due to its non-locality when searching similar pixels. In order to resolve this issue, we designed and implemented a parallel NLM algorithm based on Intel Xeon Phi hardware, which has Intel’s Many Integrate Core(MIC) architecture. The main contributions of this work are listed as follows:(1) Analyze the sequential program of non-local means algorithm, and find the main part of time consuming. Compliant with the capability of the MIC architecture, a corresponding parallel NLM algorithm is designed and implemented.(2) The parallel algorithm can achieve satisfactory speed-up, however, the obtained speedup shows a step-like distribution, i.e., the speedup changes with the inputting image sizes. In addressing this, the parallel algorithm is further optimized to deal with the preheating and load balance of multiple threads on MIC.(3) In the above-mentioned parallel algorithm, CPU is in the idle state when MIC cards are computing. In order to make full use of the computation resources, a static collaborative parallel algorithm that can use both MIC and CPU resources is developed.(4) In order to utilize the multiple MIC cards efficiently, a dynamic task schedule for the collaborative parallel algorithm is also designed. Based on the above parallel algorithm, a new parallel algorithm is designed for a 3 MIC+CPU platform that is used for this work.Finally, the experiments for the standard version, optimized version, and loadbalanced version are carried out with different sizes of input RS images. In the experiments, the results have been analyzed collaboratively with certain evaluation criteria, such as speedup, etc. Several conclusions are drawn from the experimental results.(1) the standard parallel algorithm can get better speedup with only one MIC card;(2) the optimized parallel algorithm can eliminate the step distribution of the speedup perfectly, and can also accelerate the RS image processing significantly;(3) the parallel algorithm that is based on either static or dynamic load balance strategies can further improve the performance. This work shall be found useful by researchers who are interested in either developing or applying parallel image enhancement on Intel MIC platform.