An Optimized Deep Reinforcement Learning-Based Model For Image Classification In Imbalanced Datasets

The problem of class imbalance has always been a challenge and hot topic in the field of machine learning.In the real world,the problem of non-uniformity of image data is widely present,such as in the medical field of disease diagnosis,where rare medical images are the minority class and healthy medical images are the majority class;in the field of security,the number of dangerous goods is often the minority class,while the number of other items is the majority class.In such cases,traditional classification models are prone to bias,meaning that the probability of correctly identifying the class with more samples as the correct result is higher.However,in these problems,correctly identifying minority class samples has more important application value.In recent years,reinforcement learning has become a new and emerging research direction in the field of machine learning,providing a new methodology for dealing with the classification problem of imbalanced images.In this paper,we construct an imbalanced classification model based on the deep reinforcement learning DQN method,and according to relevant research theories,we have done the following work: firstly,we improve the deep reinforcement learning imbalanced classification algorithm and propose a deep reinforcement learning algorithm(DQN+attention)that integrates self-attention mechanisms.We study the effect of channel-domain attention mechanism Se Net,ECANet,spatial-domain attention mechanism SANet,and hybrid-domain attention mechanism CBAM on the classification performance of the DQN model.Secondly,to improve the generalization ability of the integrated self-attention reinforcement learning model,we propose a feature-enhanced deep reinforcement learning algorith(Dm QNimb+SENET+SANET)that integrates ensemble learning ideas and improves the algorithm’s reward function.Using recall,F1 score,geometric mean,and balanced accuracy as evaluation indicators,we compare the performance of the proposed models with the DQN imbalanced classification model,Resnet-50,Resnet-50＿RUS,Resnet-50＿ROS,and Resnet-50＿FL on multiple imbalanced image classification datasets,including cifar-10,animal-10,and cifar-100.The research results show that integrating different types of attention mechanisms can effectively improve the model’s ability to recognize minority classes.Meanwhile,the Dm QN+SENET+SANET algorithm has certain advantages in handling class imbalance classification tasks compared to the compared models.