Research On Meta-Reinforcement Learning For Object Recognition

As a fundamental task in the field of image perception and vision,the performance of object recognition will directly affect the performance of subsequent intermediate and advanced tasks,which in turn determines the application of artificial intelligence in particular scenarios.In recent years,with the development of deep learning,great breakthroughs have been made in object recognition methods.However,the existing methods still have some difficulties in recognizing small objects in complex scenes and few-shot object lacking annotated data that have not yet been solved.In this thesis,the existing problem is analyzed from two aspects.Firstly,the object feature perception ability of deep learning and the sequence search ability of reinforcement learning are used to solve the problem of small object recognition in complex scenes.Secondly,the rapid adaptability of meta-learning is used to solve the problem of few-shot scenarios.The main work and innovation points of this thesis are as follows:1)Aiming at the problem of small object recognition,this thesis proposes a small object recognition method St-CISNet based on reinforcement learning and context information.Small objects in complex scenes face small size and low resolution,and deep convolutional networks are usually difficult to capture the deep information used to detect classes when processing these small objects,while the sequence search ability of reinforcement learning can find regions of interest in images and re-detect the selected regions.Therefore,in this thesis,reinforcement learning is introduced in small object recognition,and a reinforcement learning small object fast search method St-CISNet is proposed that integrates spatio-temporal context information.This approach differs from previous approaches to target recognition based on reinforcement learning,in that St-CISNet simulates the human eye by searching from the whole to the local,combining the temporal context search mechanism of reinforcement learning,placing the "temporal context" with historical information on top of the "positional context" with positional information.The search is gradually narrowed down to focus on a small object.The experimental results show that St-CISNet effectively improves the recognition performance of small object.2)Aiming at the problem of few-shot object recognition,this thesis proposes a few-shot object recognition method Meta-St-CISNet based on meta reinforcement learning.Existing deep learning and reinforcement learning recognition algorithms often have few-shot utilization and poor generalization performance,while the rapid adaptability of meta-learning can be quickly generalized to new tasks with few-shot.Therefore,this thesis introduces meta-learning into reinforcement learning and proposes a novel few-shot recognition method Meta-St-CISNet.This method introduces meta reinforcement learning into the field of object recognition for the first time,and realizes the rapid adaptation of new tasks under a small number of labeled samples through the two-layer optimization learning method.Specifically,Meta-St-CISNet trains a highly sensitive initialization parameter for St-CISNet,which allows the model to quickly generalize to new tasks with a small number of training shots.In addition,Meta-St-CISNet avoids the steps of data set processing in the traditional few-shot recognition paradigm,effectively simplifying the recognition process.The experimental results show that Meta-St-CISNet effectively improves the performance of few-shot object recognition.