No abstract

Removing non-uniform blur from an image is a challenging computer vision problem. Blur can be introduced in an image by various possible ways like camera shake, no proper focus, scene depth variation, etc. Each pixel can have a different level of blurriness, which needs to be removed at a pixel level. Deep Q-network was one of the first breakthroughs in the success of Deep Reinforcement Learning (DRL). However, the applications of DRL for image processing are still emerging. DRL allows the model to go straight from raw pixel input to action, so it can be extended to several image processing tasks such as removing blurriness from an image. In this paper, we have introduced the application of deep reinforcement learning with pixel-wise rewards in which each pixel belongs to a particular agent. The agents try to manipulate each pixel value by taking a sequence of appropriate action, so as to maximize the total rewards. The proposed method achieves competitive results in terms of state-of-the-art.