Methods Of Weakly Supervised Object Localization Based On Deep Reinforcement Learning

With the advent of the big data era and the application and development of deep learning technology,the performance of object detection has been rapid growth.However,the success of deep learning depends on a large amount of labeled data,and the labeling of massive data is time-consuming and laborious,and the high accuracy is difficult to ensure,which limits the application of object detection in practical tasks.In order to reduce the dependence on data annotation,researchers began to focus on WSOL(weakly supervised object localization)methods.In the task of WSOL,when only the image-level labels are known,mining the location information in the image can greatly reduce the burden of data labeling.Current weakly supervised learning methods mainly use classification models to mine location information,which causes the model to obtain inaccurate and incomplete location information.How to connect the image-level label with the location information is the key to solving the problem.For this reason,after researching and exploring related literature and methods,we use the idea of reinforcement learning to use the classification result corresponding to the image-level label as a reward to drive the agent to find better location information,and then propose the methods of weakly supervised object localization based on deep reinforcement learning,the main work is as follows:(1)We design the framework of the WSOL method based on deep reinforcement learning,which includes feature extractors and classifiers based on deep learning,and decision agents based on reinforcement learning.Aiming at the problem of object localization and weakly supervised object localization,this paper also designs a better network structure,action strategy and reward function,so that the reinforcement learning method can be combined with the WSOL methods well.According to experimental verification and analysis,in the CUB200 data set,the optimal network structure designed in this paper is specifically VGG16CONV5-GAP-4096-1024-1024-11 actions.This structure improves the stability and robustness of the reinforcement learning method while preserving image features.Based on the common 9 actions,we designed 11 actions strategy,which not only retains the basic conditions for controlling the bounding box,but also ensures the efficiency and possibility of training and testing.Our reward function uses the in-frame recognition probability and the out-of-frame recognition probability to separate the object from the background,which help us solve the problem that the classification model pays too much attention to the local features.(2)We design two object localization methods based on reinforcement learning,namely Value-based and Policy-based strategies,and experimentally verified them on WSOL tasks.At the same time,the impact of reinforcement learning parameters on performance is analyzed in the experiment,and the advantages and disadvantages of Value-based and Policy-based methods are compared.