Double DQN method for object detection

This paper proposes an object detection strategy with a deep reinforcement learning method Double DQN in which, given an image window, a deep reinforcement learning agent is trained to determine which predefined region candidates to focus the attention on. In the Double DQN framework, the first DQN is used to select an action to search the target region and the second is to evaluate the selected action. In order to verify the efficiency of our method, we compare the performance of Double DQN with the traditional DQN. Experiments indicate Double DQN has good results with higher precision and recall. The number of actions performed by the Double DQN agent are analyzed and the results show that the object can be found within very few steps. We also conducted an experiment on person detection, the results show that the algorithm has strong adaptive ability.