Research On Object Detection Based On Reinforcement Learning

Object detection is one of the hottest research directions in the field of computer vision.The goal of object detection is to locate all the presented objects in the input image or video and determine the category to which each object belongs.The typical object detection process is to cover the object with a bounding box tightly.Since the powerful feature extraction capabilities of convolutional neural networks,object detection algorithms based on deep learning have achieved tremendous improvements in accuracy and speed compared to traditional detection algorithms,however,there is still a large number of redundant regions need to be processed.The process of the regions has become a bottleneck that limits the speed of the algorithm.In order to reducing the number of regions that need to be processed,this paper proposes two object detection algorithms based on deep reinforcement learning.1.Deep reinforcement learning for visual object detection with bounding box regression.The object detection algorithm based on reinforcement learning usually adopts predefined search actions in the detection process,the shape and size of the proposal regions generated by them are not changed much,resulting in a low accuracy of object detection.For this reason,based on the deep Q-Network(DQN)object detection algorithm,we proposed an object detection algorithm by combining bounding box regression with deep reinforcement learning.Firstly,the DQN determines the search action according to the information extracted from the initial proposal regions,and then selects the next proposal region approaching the ground truth according to the action.Then repeat the above process until DQN has enough confidence to determine the current region as the ground truth,and then the search process is terminated.Finally,the current region coordinates are regressed by the regression network to achieve a better localization.The experimental results on the Pascal VOC single-category dataset show that the accuracy of visual object detection is effectively improved by the introduction of bounding box regression.2.Object detection algorithm based on multi-layer features and deep reinforcement learning.The objects of various sizes have different expression capabilities on the same depths of feature maps,using only single-layer feature maps for all object may result in poor performance in detection.Therefore,in order to fully express the features of the object of different sizes,this paper introduces multi-layer features on the basis of the object detection algorithm based on deep reinforcement learning.The reinforcement learning agent can extract the corresponding feature layer according to the region-feature mapping and combination of multi-layer features and reinforcement learning for object detection.The experimental results in the single-category object detection of Pascal VOC dataset show that the proposed algorithm can effectively improve the accuracy of detection compared with the algorithm based on deep reinforcement learning without using multi-layer information,which proved the advantages of the proposed algorithm.