Supervised Contrastive Learning-Based Long-Tailed Image Recognition

Deep learning is a branch of machine learning,which acquires the inherent laws and representation levels of data through hierarchical learning.At present,deep learning is widely used in computer vision,natural language processing,speech recognition,quantum chemistry,medicine and other fields,especially in image recognition.This is inseparable from largescale,high-quality datasets and GPU performance improvements.In traditional image recognition tasks,the class distribution of training data is often artificially influenced,that is,the number of samples of different classes is basically the same.However,datasets in practical application scenarios usually present a long-tailed distribution,which poses a great challenge for image recognition tasks,that is,deep models trained on long-tailed datasets perform poorly on tailed categories.How to use long-tailed data to train classification models that perform well in various categories has become one of the research hotspots in the field of computer vision and deep learning.This paper studies an improved method for long-tailed image classification based on deep learning.The main work is as follows:1.This paper designs a convolutional neural network loss function for long-tailed image classification.The most commonly used loss function in image classification neural networks is the cross-entropy loss.While this loss function performs well on balanced datasets,on long-tailed datasets,this loss function causes model performance to be biased towards the head category,while the tail category performs extremely poorly.To solve this problem,this paper proposes a new loss function.This loss function not only improves the representation ability of the feature extraction network during training,but also balances the performance of the classifier on each class.Specifically,the loss function not only enables the model to learn the representative features of each class of images,that is,the features of the same class are similar and the features of different classes are far apart,but also makes the model pay more attention to the tail class.2.This paper adopts the post-hoc tuning strategy of convolutional neural network for long-tailed image classification.At present,most of the datasets used in long-tailed recognition work are training sets with long-tailed distribution and validation and test sets that satisfy uniform distribution.Although the existing resampling and reweighting strategies improve the model performance to a certain extent,they do not solve the contradiction between the training phase and testing phase objectives of long-tailed learning.In order to solve this problem,according to the idea and method of optimal transport,we redistributed the labels of the network prediction results in the testing phase,so that the predicted label distribution is close to a uniform distribution.Finally,this paper evaluates the proposed method on four long-tailed datasets,and the evaluation results show that our method is more effective than most long-tail classification methods.