Design And Implementation Of Parallel Region Segmentation Algorithm On CUDA Platform

In the field of digital image processing, target objects’recognition and extraction is a common operation, whereas image segmentation remains a classic difficulty in recognition and extraction. On one hand, image segmentation needs to process large amount of data, on the other hand the complexity of this algorithm is pretty high, so CPU implementation of the algorithm can’t meet the real-time requirements. Therefore, many domestic and international scholars have carried on a great deal of research on the acceleration of region segmentation algorithms, some scholars use CUDA platform to accelerate region segmentation algorithm and have achieved good results.Considering the advantages of GPU CUDA platform for large-scale data processing and the importance of region segmentation in image processing, this paper proposed three grayscale region segmentation algorithms and their further parallel implementation on CUDA platform based on existing serial and parallel algorithms, they are Otsu, region growing and Quick shift. In order to achieve the parallelization of Otsu and Quick Shift, this paper proposed an algorithm which transforms the serial parts of those algorithms into parallel implementation by using CUDA. This method has improved the real-time performance and greatly reduced the multiple data exchange between CPU and GPU so as to achieve a better performance. As for the parallel implementation of region growing algorithm, the paper put forward a new parallel algorithm using Union-find which is different from the traditional serial algorithm. In this paper, a parallel iterative algorithm is given by using CUDA strategies, this method reduce time significantly.This paper begins with a brief explanation of some basic strategies and parallel implementation of region segmentation, and continues with details of the parallel application of different region segmentation algorithms based on CUDA platform, including parallel implementations and parallel strategies and the results of comparing with serial algorithms. The related experimental results show the superiority of CUDA technology in parallel processing, all region segmentation algorithms have been greatly improved.