Semi-supervised Image Classification Based On Feature Pyramid And Local Aggregate Coding Of Generative Adversarial Networks

The research of image classification is a basic research in the field of computer vision.Image classification is to extract and learn the features of sample images to achieve the correct classification of the class labels to which the sample images belong.Image classification tasks can be divided into fully supervised learning tasks,unsupervised learning tasks,and semi-supervised learning tasks.Fully supervised learning tasks are currently relatively mature and can achieve very good image classification accuracy,but traditional fully supervised image classification tasks have very high requirements on the training datasets,and the quality of the datasets directly determines the accuracy of the fully supervised image classification.In some tasks where it is difficult to obtain high-quality training datasets,the task of fully supervised image classification is difficult to perform or the image classification accuracy is not ideal.Although the unsupervised learning task does not have the limitation of the datasets label,the unsupervised image classification task has the problem that the image classification accuracy is not ideal.For training datasets that contain only a small number of class labels,the semi-supervised image classification model can still meet good image classification accuracy.In this paper,by building a local aggregation coding and feature pyramid semi-supervised image classification model based on generative adversarial networks,and incorporating feature enhancement strategies,we can obtain better image classification accuracy on datasets lacking labeled training samples,and reduce the image restrictions on training datasets in classification tasks.This paper studies the problem of semi-supervised image classification,and the main results are:1.This paper proposes a feature-enhanced semi-supervised image classification method based on generative adversarial networks.By constructing an auxiliary classification network to optimize the network structure of the semi-supervised classification model,the squeeze incentive architecture and feature matching items are introduced to strengthen the image category features,thereby constructing a feature-enhanced semi-supervised image classification model.The constructed model adjusts the image feature weight value through squeeze excitation,optimizes the auxiliary classification network activation function item,realizes the weighted processing of negative value feature information,and provides a richer image feature sample for the squeeze excitation module to obtain enhanced images Interclass features,improve the accuracy of semi-supervised image classification;the constructed model optimizes the loss function of the auxiliary classification network and the generation network by introducing regularization terms,alleviating the over-fitting phenomenon of the auxiliary classification network and enhancing the target gradient of the generation network,the quality of the fake image is improved;in the process of semi-supervised network training,the artificial classification is added to the auxiliary classification network and the label of the training datasets is adopted to take a one-sided label smoothing operation to improve the gradient of the auxiliary classification network.The feature-enhanced semi-supervised classification model improves the boundary recognition between image classes by enhancing inter-class features,and obtains good image classification accuracy in semi-supervised image classification tasks.2.This paper proposes a feature pyramid semi-supervised image classification algorithm based on generating an adversarial network.In order to enhance the diversity of features between image classes,a pyramid-shaped feature graph sub-structure is introduced to construct a feature pyramid semi-supervised image classification model to achieve the fusion of high-level semantic features and low-level features of multi-layer spatial images.The constructed model extracts the semantic features of different layers for the auxiliary classification network,the shallow network extracts the basic features such as the texture features and color features of the pyramid-shaped sub-images,and the deep network extracts the pyramid-shaped sub-images with high-level shapes and contours feature,and then merge multi-layer spatial sub-features to construct diverse image inter-category features,highlight the difference between the features of different types of images,and clarify the boundary between different types of images.The experimental results show that the proposed algorithm effectively improves the accuracy rate of semi-supervised image classification is discussed.3.This paper proposes a local aggregation coding semi-supervised image classification method based on the generation of adversarial networks.In order to realize the selection and optimization of features between image classes,a local aggregation coding model is introduced,and the local aggregation coding training method is optimized,and a semisupervised image classification model of local aggregation coding is constructed to improve the recognition of feature differences between image classes.The constructed model extracts the sample image features through the auxiliary classification network,and then performs local aggregation coding on the extracted sample image features,and establish the mapping relationship between the features of the sample image and the local aggregation codebook through network inference,and obtain the visual descriptor of the image feature after the local aggregation coding.Use the visual descriptor of the image feature after the local aggregation coding to perform semi-supervised image classification.The proposed method achieves the optimization of image class features through local aggregation coding,can extract image class features with sparse characteristics,improves the image class boundary recognition,and obtains high classification accuracy in semi-supervised image classification tasks.