Multi-threshold Image Segmentation Techniques Based On Intelligent Optimization Algorithm And Its Parallel Speedup

Image segmentation is one of the key technologies in the field of digital image processing and computer vision, and it’s also the basic work in image recognition and image analysis. Its result will directly affect the correctness of the subsequent operation. In the existing image segmentation algorithms, threshold method is widely applied because of its stable performance and implementation simplicity.In an image, while the object is clearly distinguishable from the background, its histogram will appear double peak. For this kind of image, bi-level thresholding could obtain a satisfied result. However, in the real world, the histograms of images are always multi-modal that much more complicated than double peak ones. For this case, multi-level thresholding is much more appropriate. The traditional threshold methods search for the optimal threshold based on a particular optimization function in the range of the gray level which has high computational complexity. At present, intelligent optimization algorithms are frequently-used for such problems. While dealing with a complicated nonlinear optimization problem, the intelligent optimization algorithm cannot assure the global optimization result, but it can get the result in a short time, so it is suitable for image multi-level thresholding.The gravitational search algorithm (GSA) is a kind of new intelligent optimization algorithm, which has high global search ability and fast convergence speed. In this paper, GSA is introduced into optimal multi-level thresholding and two kinds of improvement was put forward:(1) As the threshold value is integers, a new gravity coefficient update mode is proposed to avoid redundant precision arithmetic and improves the convergence speed.(2) Mutation in optimal individual, which avoids premature convergence and further improves the solution quality. The results of the simulation show that the proposed algorithm can quickly converge to global optimal solution, the overall performance is better than contrast algorithms.In order to further improve the real-time performance of the proposed algorithm, CPU multi-thread technology and CUDA programming model are used to parallel accelerate the gravitational search algorithm and the maximum entropy method respectively. The simulation experiments show that in the same experimental conditions the improved method could obtain high speedup ratio as240%and meet the real-time requirements.