一种基于Q学习的图像阈值确定方法

摘要

The threshold value of the image in the image processing is very important.The histogram method is a commonly used threshold value determination method,but not to determine the threshold value of the non-bimodal image.Reinforcement learning is learning through interaction with the environment.Q-learning is to strengthen one of the main methods of learning.This article describes a method of using the Q-learning algorithm to determine the optimal threshold.Agent is from a constant threshold algorithm and applied to the image.In the case of the objective,the return is defined based on the ratio of black pixels,the object region,the deviation of the tolerance area,the number of objects.Agent maps the state of the environment to the appropriate action,and tries to get the best return.The experiments show that the proposed method can be objective or subjective way to integrate the expertise of the people,in order to overcome the inadequacies of the existing methods.%图像的阈值在图像处理中非常重要.直方图法是常用的阈值确定方法,但无法很好地确定非双峰图像的阈值.强化学习是通过与环境的交互来学习,Q学习是强化学习的一种主要的方法.本文介绍一种使用Q学习算法确定最优阈值的方法.在该算法中,Agent从一个恒定的阈值开始,并把它应用到图像.在客观的情况下,回报是在黑色像素的比率、对象区域、公差面积的偏差、对象的数量的基础上被定义的.Agent将环境状态映射到适当的动作,并尝试获得最大回报.实验表明,所提出的方法可以用客观或主观的方式整合人的专业知识,以克服现有方法的不足之处.