Reinforcement Learning-Based Image Exposure Reconstruction for Homography Estimation

The homography matrix plays a vital role in robotics and computer vision applications, but mainstream estimators are usually customized for specific problems and are sensitive to image quality. In response to this situation, a reinforced agent is proposed to improve image quality by sequentially reconstructing the exposure. First, the gamma correction theory is employed to design a nonlinear exposure adjustment function so that the agent’s action is not bound to additional hardware or software. Then, the agent is designed as consisting of a metric network and a Q network that are trained under the reinforcement learning framework. When a black-box nondifferentiable homography estimator is given, the metric network can map the image into its corresponding embedding space, and the Q network can further determine an exposure value to produce pleasing images for it. Comprehensive experiments are conducted on homography samples generated from the public aerial DOTASet. After reconstructing the exposure of the original input, all selected estimators can obtain more accurate results. It also reveals that visually satisfactory images may not always be the best choice for homography estimation.