In many vision-based applications we need to segment an object of interest in digital images. For methods which rely on a learning process, the lack of sufficient number of training samples is usually an obstacle, especially when the samples need to be manually prepared by an expert. In addition, none of the existing methods uses online feedback from the user in order to evaluate the generated results and continuously improve them. Considering these factors, a new algorithm based on reinforcement learning is discussed in this chapter. The approach starts with a limited number of training samples and improves its performance in the course of time.
A potential obstacle when we apply reinforcement learning into image-based applications is the large number of state-action pairs. In such cases, it is usually difficult to evaluate the state-action information especially when the agent is in the exploration mode. The opposition-based leaning is one of the methods that can be applied to converge to a solution faster in spite of a large state space. Using opposite actions we can update the agent’s knowledge more rapidly. The experiments show the results for a medical application.