Semi-Supervised Image Classification Method Based On Neighborhood Feature Fusion

Image classification is an important research direction in the field of computer vision.Traditional image classification requires sufficient labeled data to ensure the performance of the classifier.However,in some practical applications,obtaining labeled samples can be difficult.Semi-supervised classification addresses the problem of weak generalization of classifiers when labeled samples are scarce by using a large number of unlabeled samples with only a small amount of labeled samples.In recent years,the application of deep learning in semi-supervised image classification has greatly improved the accuracy of semi-supervised image classification tasks.Among the current semi-supervised image classification methods,consistency regularization has shown outstanding performance.The main idea of such methods is to encourage the model to make consistent predictions for the same image under different transformations that preserve semantics,thus enhancing the model’s generalization ability.However,existing consistency regularization methods have some limitations.First,the transformation methods used in consistency regularization usually only operate in the input image space,without considering different transformations in the feature space.Transformations in the feature space can generate more diverse semantic features.Second,consistency regularization methods often overlook rich neighborhood information during the feature learning process.The neighborhood information of instance points can help the model learn more discriminative features.In this paper,we propose a semi-supervised image classification model based on neighborhood feature fusion to address the problems existing in the aforementioned consistency regularization methods.Thanks very much for these comments.We have added the missing citation numbers in the main text and ensured they are in the correct sequences.This article proposes a semi-supervised image classification model based on the backbone convolutional neural network,which incorporates a nearest neighbor graph and introduces a feature fusion module based on multi-head attention.Firstly,a memory bank mechanism is employed to retain the labels and feature information of all samples during the training process,enabling accurate retrieval of nearest neighbor samples for each training sample from a larger pool of training samples.To preserve historical information and obtain more accurate features and reliable pseudo-label predictions,a memory bank update mechanism based on temporal ensembling is designed.After each mini-batch sample training,an exponential moving average method is used to integrate the saved features and pseudolabel information from previous epochs and update the memory bank.Then,the similarity between all sample features in the memory bank is used to determine the nearest neighbor nodes for each sample,constructing a nearest neighbor graph.A label consistency regularization term based on the nearest neighbor graph is proposed,aiming to align the label distributions of target nodes with those of neighboring nodes as much as possible.Furthermore,a neighborhood feature fusion module based on multi-head attention is introduced to enhance the target features by fusing the features of neighboring samples.This module utilizes the multi-head attention mechanism to dynamically adjust the fusion weights of each neighboring feature during the fusion process,assigning greater weights to important neighboring features and bringing similar sample points closer together in the feature space.Finally,the proposed semi-supervised image classification model is evaluated on the MNIST,SVHN,and CIFAR-10 datasets,and the experimental results confirm the effectiveness of the new model.