SA-PSO算法在CNN边缘提取模板设计中的应用

提出了一种用于边缘提取的细胞神经网络（CNN）模板的设计方法，该方法在基本粒子群算法的基础上引入模拟退火机制，形成模拟退火粒子群算法（SA-PSO）对模板参数值进行搜寻。在搜索过程中，用退火温度调节粒子的突跳概率，轮盘赌策略确定粒子的全局最优的替代值，这样能有效避免基本PSO算法容易陷入局部最优解的问题。同时，为了保证每轮搜寻产生的解均能使CNN网络稳定，用CNN反馈模板的研究结论对粒子群解空间进行约束。模拟实验表明，文章算法设计出的CNN模板有良好的边缘提取能力。

The application of SA-PSO to CNN edge detection template design

In this paper, a synthesis procedure of cellular neural networks (CNNs) template design for edge detection is proposed. This method combines simulated annealing and particle swarm optimization (SA-PSO) to search template values. In the process of searching, annealing temperature is used to adjust kick probability and roulette is adopted to select global optimal replacement value. In this way, trapping into local optimum problems of PSO can be effectively avoided. Moreover, in order to guarantee stable outputs of CNNs, properties of CNN feedback template obtained in previous researches are used to constraint particle solution space. Simulation results show CNN templates designed by this method are efficiency in edge detection.