Context-based Fine Hierarchical Object Detection with Deep Reinforcement Learning
Yihong Cao, Xiao Bo Lu, Yiwei Zhu, Xuanyu Zhou · 2020
We present a method for object detection by adopting an agent that follows a fine hierarchical search strategy. The key idea of this strategy is that the agent detects the target on region proposals by analyzing the image information of the search region at each level. Start from the entire image, the agent detects the subregion that contains the object level by level until it confirms the target has been found. We model this process as a Markov Decision Process (MDP) to use reinforcement learning to train the agent to learn the optimal strategy. By utilizing a hierarchical search strategy, we can avoid the generation of a large number of region proposals. Besides, we add the ReNet layer that extracts context information to improve the agent’s ability to understand the information conveyed by the image. The experiments we conducted have demonstrated the efficiency of our proposed method. Experiments demonstrate the efficiency of our proposed method, the agent only needs 2 or 3 steps to detect the target in images under most circumstances.