Target Detection In Reinforcement Learning Based On U-shaped Network

Object detection is to locate and determine the category of the object in an image or video frame.With the development of convolution neural network,target detection technology has achieved a qualitative leap from traditional to advanced,The target detection algorithm based on deep learning uses CNN as the basic framework to automatically learn high-level features from the input data layer by layer,and uses the network layer of CNN to complete feature extraction,feature selection,classification and other work,and finally realizes the target detection and other tasks.The researchers have been improving on the basis of them,and put forward a series of more perfect detection algorithms,which have made great progress in accuracy and speed.However,even so,there are still some shortcomings,that is,the "small object" is easy to miss detection and the "detection box" position is still more serious,Compared with large object,small object detection has always been a big problem in the field of image processing because of its low pixel,less proportion relative to the whole image,large interference from the image background and easy occlusion.To solve the above problems,this thesis proposes a target detection algorithm based on deep reinforcement learning to improve the feature attention of small targets and the accuracy of the whole algorithm.The main work is as follows:Firstly,aiming at the problems of small target and occlusion caused by small pixel,large background interference and occlusion,the attention enhancement learning algorithm based on improved u-network is proposed to improve the small target,In the feature extraction stage,the U-shaped network framework is used to extract features,and the multi-dimensional attention module is embedded in the horizontal connection.The attention mechanism is used to speculate and determine the possible features around the target,and the U-shaped network is assisted to improve the feature exploration ability of small targets and occluded targets.Finally,the final feature is obtained by vector stitching.Secondly,in the target detection stage,because the regression operation of deep learning network can not tightly surround the target,it will lead to more useless pixels in the target area,which makes the detection accuracy is not high.Therefore,based on the improved U-shaped network,this thesis introduces reinforcement learning network to improve the detection accuracy.In this part,two kinds of reinforcement learning networks are selected to evaluate the algorithm,which are traditional Deep Q Network(DQN)and Soft Actor Critic networks(SAC).In order to improve the detection accuracy,the action of agent selection is defined as the action of deforming and shifting the detection frame,so that the detection frame tightly surrounds the target.Through the final adjustment of the detection frame in the reinforcement learning part,the detection accuracy is improved.This experiment is carried out on the standard target detection data set Pascal VOC 2007 and Pascal VOC 2012,and the average accuracy is 79.4in a single category of target data.The experimental results show that the proposed reinforcement learning target detection algorithm based on improved u-network can achieve better detection effect than other target detection algorithms.