A Study On Swarm Intelligent Algorithms For Multilevel Thresholding

As a segmentation method with strong practicability, the degree of practicability of thresholding does have a direct effect on the efficiency of the whole processing system, therefore, the performance of the practicability of thresholding becomes an important index to measure its superiority.With the increase of threshoding numbers, traditional thresholding methods can no longer meet the requirements of real-time applications. Currently, combining swarm intelligent algorithms which have strong optimization capability to find optimal thresholds based on certain criterion becomes a hot research spot.After analysing the mechanism of seven swarm intelligent algorithms, comparison stduies of the seven algorithms with respect to the optimization precision, stability, convergence speed and success searching ratio in the multilevel thresholding based on OTSU method and Kapur entropy method have been made. Results of the experiments on the test images indicate that the performance of swarm intelligent algorithms is different when different segment criterions used. But considering the searching precision, stability, convergence speed and success searching ratio, the Cuckoo Search have a superior performance on multilevel thresholding based on both the OTSU method and the Kapur entropy method, while Artificial Bee Colony Algorithm and Shuffled Frog Leaping Algorithm have an inferior performance on both methods when compared to other algorithms.After analysing the mechanism of several swarm intelligent algorithms and pattern search algorithm, a strategy which related to pattern search algorithm has been proposed to improve the swarm intelligent algorithms' performance on multilevel thresholding. The strategy is that applying the pattern search algorithm with fixed stepsize to the global history best solution of each iteration of the swarm intelligent algorithms. Results of the experiments indicate that the mechanisms of swarm intelligent algorithms which have a better balance of global exploration and local exploitation would lead to better improvements after the proposed strategy applied. But the strategy has different degree of improvements in the performance of swarm intelligent algorithms when different specific thresholding criterions used. When OTSU method is used, the strategy can improve the performance of most algorithms when the number of thresholds is small, and can also improve the peformance of most algorithms on which the image histogram feature is smoothing when the number of thresholds is large. When Kapur entropy method is used, the strategy can improve the searching precision, stability and convergence speed of Shuffled Frog Leaping Algorithm, Cat Swarm Optimizaiton, Cuckoo Search, and improve the optimization precision of Artificial Bee Colony Algorithm and Firefly Algorithm, but the improvements of Particle Swarm Optimization and Bat Algorithm are unsatisfactory.