Reinforced Segmentation of Images Containing One Object of Interest
MetadataShow full item record
In many image-processing applications, one object of interest must be segmented. The techniques used for segmentation vary depending on the particular situation and the specifications of the problem at hand. In methods that rely on a learning process, the lack of a sufficient number of training samples is usually an obstacle, especially when the samples need to be manually prepared by an expert. The performance of some other methods may suffer from frequent user interactions to determine the critical segmentation parameters. Also, none of the existing approaches use online (permanent) feedback, from the user, in order to evaluate the generated results. Considering the above factors, a new multi-stage image segmentation system, based on Reinforcement Learning (RL) is introduced as the main contribution of this research. In this system, the RL agent takes specific actions, such as changing the tasks parameters, to modify the quality of the segmented image. The approach starts with a limited number of training samples and improves its performance in the course of time. In this system, the expert knowledge is continuously incorporated to increase the segmentation capabilities of the method. Learning occurs based on interactions with an offline simulation environment, and later online through interactions with the user. The offline mode is performed using a limited number of manually segmented samples, to provide the segmentation agent with basic information about the application domain. After this mode, the agent can choose the appropriate parameter values for different processing tasks, based on its accumulated knowledge. The online mode, consequently, guarantees that the system is continuously training and can increase its accuracy, the more the user works with it. During this mode, the agent captures the user preferences and learns how it must change the segmentation parameters, so that the best result is achieved. By using these two learning modes, the RL agent allows us to optimally recognize the decisive parameters for the entire segmentation process.