Research On Long-Tail Classification Based On Causal Inference

In the field of computer vision,the emergence and development of deep neural networks has led to an increasing range of related applications.For example,data segmentation,classification and detection tasks of images and videos by training deep neural networks have surpassed humans in terms of accuracy and speed.However,the ideal computer vision training environment and the actual application scenario generally differ to a greater or lesser extent.One of the major differences is the data distribution in reality versus the data distribution in computer vision research,where most of the data in the actual application scenario has a long-tailed distribution,i.e.,a few categories contain a large number of samples and most categories contain a small number of samples.Since most of the current network models use a data-driven approach,the long-tailed distribution of data tends to cause a decrease in the recognition accuracy of network models.In addition.most of the deep neural network models are still "black box" models.and it is difficult to explain the final decisions made by the models.Although traditional long-tailed classification models can obtain better classification results by rebalancing the data.their interpretability is poor,and the spurious association between irrelevant features and category labels in the classification process makes the model less stable in prediction,which affects the final classification accuracy of the model.The main research content of this paper is summarized as the following three points:(1)To address the problems of poor interpretability and unstable prediction of traditional long-tailed classification models.this paper proposes a long-tailed classification model based on causal inference,which combines the theory of causal inference to improve the interpretability of the model,and removes spurious correlation by reweighting to improve the stability of the model when predicting.And for the drawback that the de-correlated sample reweighting method generates a large amount of extra space in the training process,this paper improves it by adding an iterative optimization mechanism in the tail to save only the optimal weights,while adding the Mixup data enhancement method in the head of the model to further improve the model performance.The overall performance of the method is proved to be superior by comparison tests with the traditional method.(2)In order to address the problem that the causal inference-based long-tailed classification model has poor recognition accuracy in the tail category containing a small number of samples when trained on unbalanced datasets,this paper further proposes its optimization model,a two-stage causal inferencebased long-tailed classification model.The model is trained in two stages,the first stage is the causal inference-based long-tailed classification model.and the second stage is the balancing fine-tuning stage to improve the recognition accuracy of the tail category by rebalancing the data set before training.In addition.a CAM-based class balancing sampling method is used in the balancing fine-tuning phase to improve the interpretability of the model at the data level.In this paper,extensive experiments are conducted on two datasets.CIFAR10_LT and CIFAR100_LT,and it is demonstrated that the performance of the model has been improved after adding fine-tuning experiments.Compared with other long-tailed classification models,the model in this paper has obvious advantages.(3)Based on the two-stage long-tailed classification model using causal inference,this paper designs a long-tailed image classification system,which includes four modules:model selection module,data import module,classification module,and result display module.Through this system,the model proposed in this paper can be applied in practice,thereby increasing its practical value.