Research On Image Classification Algorithm Based On Imbalanced Samples

When the era of intelligent comes,deep learning technology has achieved excellent results in multi-disciplinary fields,but deep convolutional neural networks rely on a large amount of high-quality labeled data when solving image classification tasks.Natural environmental factors and human factors during image acquisition and labeling will adversely affect the performance of the model,especially class-imbalanced and difficulty-imbalanced training set will cause a network containing a large number of neurons tends to learn a simple mapping,which reduces accuracy when inferring online.Therefore,this paper aims to solve the low classification accuracy problem caused by imbalanced training set,and proposes a convolution network structure that integrates attention,a two-branch network model based on difficult sample mining,and a minority sample generation strategy based on generative adversarial networks.Finally,the purpose of improving the accuracy and recall of image classification is achieved.The specific work of this article is as follows:(1)Aiming at the problem of classification of difficult samples affected by complex backgrounds and local differences,a convolutional network structure that combines space domain and channel domain attention is proposed.This method firstly uses asymmetric pooling and 1 × 1 convolution to increase the detail of attention information.Secondly,it uses a learnable soft attention mask and self-attention to activate convolution feature maps.This model not only focus on some features,but also captures the long-range dependence between pixels,which effectively solves the problem of difficult sample classification.(2)Aiming at the imbalance problem,a two-branch network model based on online difficult sample mining is proposed.The model firstly calculates the loss of training samples from the base classification network,and then uses the isolated forest algorithm to divide the samples,transfer samples to a random depth sub-network or normalized level weighted subnetwork accordingly to complete the training.This model use the cross-entropy loss and result integration to complete the classification training.What's more,the effective sample weighting loss can effectively solve the class-imbalanced problem.(3)Aiming at the extremely class-imbalanced problem in data sets,based on the strategy of generating minority samples,a generation adversarial network model with a sub-gradient penalty and relaxation term is proposed.By adding a layer normalization strategy in the generator network,the diversity and image quality of the generated samples are improved.Finally,it is proved that combining the generated samples and the original training samples can alleviate the imbalance between classes and improve the classification accuracy.