Research On Class Imbalanced Image Classification Algorithm Based On Sample Mixing Technolog

With the rapid development of deep learning,powerful deep learning models have been widely used in various fields,such as computer vision,natural language processing,intelligent speech,etc.In the application scenarios based on deep learning technology,large data sets are essential to ensure the performance of deep learning models.However,the real data distribution often has an imbalanced distribution of categories.In the training process,categories with few samples are easy to be ignored by the model,which damages the performance of the model and leads to the performance of the trained model falling short of expectations.Therefore,how to use training samples with imbalanced category distribution to effectively learn has become a research direction that attracts much attention and needs to be solved urgently in recent years.Based on image classification tasks,this paper studies the image classification problem of imbalanced category.The relevant solutions are discussed from the data level of Mixup data enhancement technology,and we propose an image classification method including Mixup combined resampling and an image classification method based on image clipping mixing.This paper proposes an image classification method based on Mixup joint resampling.By analyzing the complementarity between oversampling and Mixup data enhancement techniques,we further propose a mixed sample method combining instance-based sampling and class-based sampling to solve the image classification of imbalanced category.In addition,this technique proposes a hybrid weight attenuation strategy,which adaptively adjusts the mixing proportion between two samplers according to the number of training epochs,so as to guide the learning of the model.Under the influence of this strategy,the weight of instance-based sampler is larger in the early stage of training,and the model is guided to establish the basic representation ability.With the progress of training,the mixing weight is gradually dominated by class-based sampler.On the basis of certain representation ability,the model is further directed to focus on the learning of rare categories.The proposed method effectively combines two sampling methods with different properties through a weight attenuation strategy,and fully utilizes the advantages of the two sampling methods to train the model.By adjusting mixing weight,the model’s attention is transferred from representation learning to rebalancing learning,and the classification accuracy of imbalanced category is significantly improved.Subsequently,this paper continues to conduct in-depth discussion on Mixup sample mixing technology.From the perspective of mixing mode and sample distribution,we propose an image classification method of imbalanced category based on image clipping.The method consists of image block mixing module and training sample selection mechanism module respectively.Image block mixing uses rectangular blocks between images to cut and paste instead of the original linear weighted mixing.This method believes that the image block mixing can make the model learn the local features between image blocks more clearly and solve the problem of feature ambiguity caused by linear weighted mixing.In addition,an adaptive box position adjustment method for image clipping is proposed to ensure the accuracy of clipping area.According to the proportion of image blocks in the mixed samples,the training sample selection mechanism selectively replaces part of the mixed samples and introduces the original samples into the mixed samples for training.Finally,the proposed method is integrated with common resampling and reweighting techniques to further improve the image classification accuracy of imbalanced category.The main contributions of this paper are as follows:(1)This paper first focuses on the limitations of Mixup data augmentation technology in the case of imbalanced distribution,and then proposes a joint resampling method based on Mixup to effectively reduce the generation of noise samples in the mixing process;At the same time,a hybrid weight attenuation strategy is proposed to guide the model training.It only needs to adjust the distribution of mixed samples to guide the model gradually focus on the features of rare classes.(2)This dissertation analyzes the performance differences between linear interpolation and image block.Firstly,a patchbased method is proposed to classify imbalanced images,which can significantly improve the feature representation of mixed samples;Secondly,the method is effectively integrated with classical resampling and reweighting techniques to further improve the image classification performance of the model in the scenario of imbalance category.