| Deep semi-supervised learning is to use a small amount of labeled data and a large amount of unlabeled data to automatically learn to the deep feature representation of the data,thus improving the generalization ability of the model on unlabeled data.Deep semisupervised learning has been widely used in the fields of image,sound,and natural language,and has become one of the important hotspots of research in deep learning.Since labeled data is costly,labeled data can be increased by data augmentation techniques,and data augmentation can be used to improve deep semi-supervised learning performance.Data augmentation is to generate new training data by performing a series of transformation operations on the original data.However,this data augmentation method still has the following limitations.(1)The effectiveness of data augmentation methods heavily depends on the similarity between unlabeled and labeled data,and when the distribution of unlabeled and labeled data is widely different,the mismatch of the augmented data will lead to the degradation of model performance.(2)The use of data augmentation techniques to extend the dataset will have the problem of semantic information loss,which misleads the model to learn invalid feature representations and thus reduces the discriminative ability of the model.Based on the above problems,the following three works are done in this paper.1.A label propagation contrast semi-supervised algorithm is proposed,which first labels samples in the unlabeled dataset by automatically labeling them,and then expands the labeled data into a larger dataset with high reliability.Then use the contrast learning method to pretrain samples with similar features to pull close,and propagate the labels from the labeled data to the unlabeled dataset using label propagation on the dataset,and finally use the similarity alignment method to select samples with high similarity from the propagated labels using a similarity threshold to construct a reliable dataset.We apply the algorithm to natural image datasets for testing,and the experimental results show that the algorithm can reduce the effect of data distribution mismatch on the semi-supervised algorithm and improve the accuracy of classification.2.An interpolation-contrast semi-supervised algorithm is proposed,which first augments the data,then performs an interpolation operation on the normalized low-dimensional representation of the sample of unlabeled data,and then also performs an interpolation operation in the feature space to extract image information from both aspects and generate a contrast loss,so that the positive sample constructed by interpolation maintains the original semantic information.We apply the algorithm to a natural image dataset for testing,and the experimental results show that the algorithm improves the discriminative ability of the network in the case of label semantic loss and makes the classification results more accurate.3.The implementation of an image classification visualization system that is capable of selecting the aforementioned label propagation contrast semi-supervised algorithm and interpolation contrast semi-supervised algorithm for quantitative evaluation of the selected images in terms of category.This section details the system requirement design,the system framework structure,and the system implementation.The classification system consists of three modules:feature extraction visualization,where a network-processed feature map is obtained after uploading an image;image classification visualization,where an image is selected for classification from under a local path,and two algorithms are selected to qualitatively analyze the categories of the uploaded image and display the classification results of the algorithms;and classification statistics display,which directly The number of each category of uploaded data is counted to understand the performance of the model intuitively. |
