Study On Improved PSO With Multi-Strategy And Its Application On Multi-Threshold Image Segmentation

Particle swarm optimization(PSO)algorithm is a typical swarm intelligence algorithm.Compared with other swarm intelligence algorithms,PSO has fewer parameters and faster convergence speed,and it is easier to carry out.However,PSO is also prone to local optimums.Aiming at this,two improved algorithms are proposed.The proposed algorithms are studied on their capabilities with some benchmark functions and they are applied to multi-level image segmentation.Experiments are carried out along with original algorithms.The main contribution of this study is as follows.First,to reduce the negative influence of the overemphasis of gbest,the dimensional information of particle is introduced to be a new example.The new information source can not only improve swarm diversity but also promote the communication between particles' dimensions.Time hierarchy or iteration adaptive strategy is extended from probability hierarchy in terms of particle's update equation.The thought of simple PSO is adopted to improve algorithm's convergence speed.Thus,two improved algorithms probability hierarchy simple PSO(PHSPSO)and time hierarchy simple PSO(THSPSO)are proposed.Experiments are conducted on fifteen benchmark functions.The results demonstrate the two proposed algorithms both have excellent performances for basic functions compared with other popular PSO variants.Probability hierarchy strategy is more effective than others in general.Second,since PHSPSO and THSPSO still have room for their search performance on complex benchmark functions,another improved PSO algorithm named GeESPSO is proposed.The proposed algorithm combines geese flight PSO and extended PSO,which moderates over convergence of particles and make full use of other particles' information at the same time.This adopted strategy may slow down the convergence speed of algorithm,thus,simple PSO is combined with it to reduce the negative influence.Finally,extended simple PSO based on geese flight(GeESPSO)is proposed.Fifteen benchmark functions are conducted to evaluate algorithms.Experimental results demonstrate that GeESPSO performs best on basic functions compared with other algorithms and it is also better than PHSPSO and THSPSO.GeEPSO performs better than GeESPSO in terms of complex functions,which accords with “No free lunch theorem”.Finally,the improved PSO algorithms are combined with max between-class variance method.Segmentation results demonstrate that the improved algorithm PHSPSO is the best in case of between-class index while GeESPSO is the best in terms of similarity index.On the whole,the improved algorithms are effective in image segmentation.