Research On Unbalanced Data Classification Based On Generative Adversarial Network

With the significant improvement of deep learning methods in computer vision,natural language processing,and other fields,people have gradually realized the significance of massive data to models.However,massive data of-ten show significant imbalances,and this data imbalance problem has become an important factor restricting the subsequent development of deep learning models to a certain extent.Traditional data processing and model optimization methods are difficult to cope with various imbalanced data and disparate imbalanced ratios encountered by deep neural network models in solving complex problems.At the same time,the significant breakthroughs made by generative adversarial networks in the field of image generation have made researchers realize the great potential of generative adversarial networks in generating data.Therefore,combining the generative capabilities of the generative adversarial network to improve the performance of the model in the case of imbalanced data has become a subject worthy of study.In this paper,we focuses on the imbalance problem in visual relationship detection and typhoon intensity estimation tasks,and explores the role of gener-ative adversarial networks in solving the imbalance problem.And we proposes a feature activation method based on generative adversarial networks,which uses the semantic and spatial information between objects in a visual relationship to activate visual features to improve the relationship of visual information and uses the adversarial network to generate activation weights to alleviate the problem of insufficient training caused by imbalanced data in relation detection tasks.Finally,multiple experimental results on the VRD dataset show that the results of our model in the Zero-shot subtask exceed the best results in the dataset,proving the effectiveness of the generative adversarial network in solving the imbalance problem in visual relationship detection.At the same time,we proposes a context-aware CycleGAN model which learns the evolution features of typhoons between different intensities to generate the typhoons visual feature of specific intensities,to achieve the goal of balancing the number of typhoon samples of different intensities,and solving the unbalanced problem of typhoon intensity estimation.We perform related experiments in the cyclone dataset and verify the effectiveness of the algorithm.