群智能算法优化的结合熵的最大类间方差法与脉冲耦合神经网络融合的图像分割算法

Image segmentation algorithm based on fusion of group intelligent algorithm optimized OTSU-entropy and pulse coupled neural network

针对最大类间方差准则下的图像分割结果携带原图信息量不足、实时性差和脉冲耦合神经网络(PCNN)模型中循环迭代次数难以确定的问题,提出了群智能算法优化的结合熵的最大类间方差法(OTSU-H)与PCNN融合的自动图像分割算法。首先,充分利用图像的灰度分布信息和相关信息,将图像信息中冗余度、竞争性以及互补性有效地融合,构造二维和三维观测空间,提出了OTSU-H准则的快速递归算法;其次,将快速递推算法的目标函数分别作为布谷鸟搜索(CS)算法、萤火虫算法(FA)、粒子群优化(PSO)算法和遗传算法(GA)四种群智能算法的适应度函数;最后,将优化之后的OTSU-H引入PCNN模型中自动获取循环迭代次数。实验结果表明,与原始的最大类间方差法(OTSU)、最大熵准则以及基于图论分割、像素的聚类分割和候选区域语义分割的图像分割算法相比,所提算法具有较好的图像分割效果,同时降低了计算复杂度,节约了计算机的存储空间,具有较强的抗噪能力。所提算法时间损耗少、不需要训练的特性使得算法的运用范围较广。

The image segmentation results under the maximum interclass variance criterion have the problems that the original information is not enough, the real-time performance is poor, the number of iterations in the Pulse Coupled Neural Network （PCNN） mode] is difficult to determine. In order to solve the problems, a new automatic image segmentation algorithm was proposed based on the fusion of group intelligent algorithm optimized OTSU-entropy （OTSU-H） and PCNN. Firefly, the gray distribution information and related information of the image were used to fuse redundancy, competition and complementarity of the image effectively, at the same time, the two-dimensional and three-dimensional observation space were constructed. The fast recursive algorithm of OTSU-H criterion was proposed. Secondly, the objective function of the fast recursive algorithm was respectively used as the fitness function of the four group intelligent algorithms of Cuckoo Search （CS） algorithm, Firefly Algorithm （FA）, Particle Swarm Optimization （PSO） algorithm and Genetic Algorithm （GA）. Finally, the optimized OTSU-H was introduced into the PCNN model to acquire the number of iterations automatically. The experimental results show that, compared with the original OTSU, the maximum entropy criterion, the image segmentation algorithms based on graph theory segmentation, pixe] clustering segmentation and candidate region semantic segmentation, the proposed algorithm has better image segmentation effect, reduces the computational complexity, saves the storage space of the computer, and has strong anti-noise ability. In addition, the proposed algorithm has a wide range of applications with the characteristics of less time consumption and not need training.