Image Enhancement Using Reinforcement Learning for Underwater Images

Conventional image enhancement techniques are based on preset algorithms and are not flexible enough to handle a wide range of image attributes. Depending on the type and content of an image, several techniques and models need to be tested and trained until one technique that works specifically for that image set is revealed. Not only is this process extremely time-consuming, but it also rules out the possibility of testing how the mixing and matching of such algorithms can lead to better results. This research investigates reinforcement learning, specifically Policy Gradient and Q-learning, to teach an agent the best enhancement technique and uncover the sequence of actions that enhance the image in the best way possible. The paper presents an Agent class that moves through states determined by the number of features recognized while interacting with an image environment. To prioritize actions that result in the most favorable feature changes, the system is built to learn from interactions with the image environment and update its rewards and value functions accordingly.