Maximum entropy-based optimal threshold selection using deterministic reinforcement learning with controlled randomization

Abstract

Traditional maximum entropy-based thresholding methods are very popular and efficient in the case of bilevel thresholding. But they are very computationally expensive when extended to multilevel thresholding since the inevitable exhaustive search of optimal thresholds needed to maximize the posterior entropy. In this paper, a reinforcement learning (RL) approach is proposed for the maximum entropy thresholding. We show that finding the optimal thresholds using the maximum entropy criterion is equivalent to learning an optimal policy of the RL problem. Therefore, the powerful Q-learning algorithm, which is widely used in RL, can be employed to eradicate the computation burden of the maximum entropy-based thresholding methods. The experimental results show that the proposed method is suitable in the case of multilevel thresholding and the performance is better than that of the genetic algorithm-based entropy thresholding method.