Semi-supervised image classification algorithm based on double threshold FixMatch

Abstract

To address the problem of how to improve the utilisation of a large number of unlabelled samples and the accuracy of classification using a few labelled ship image samples, this paper proposes a semi-supervised ship image classification network based on Double threshold FixMatch (DT-FixMatch). Firstly, double threshold is introduced into the FixMatch, and the model predicts the unlabelled images after weak enhancement, and retains the category of the results with a certain confidence level higher than the high threshold and transforms them into pseudo-labels; for the results higher than the low threshold, the softmax output value of the prediction result is directly used to compare it with the model's softmax output value of the strongly-enhanced version of the same image, and calculates the Mean Square Error (MSE) loss, which allows the low confidence samples to be fully utilised. Secondly, to reduce the influence of noisy pseudo-labels, we have implemented label smoothing and consistency regularization. Additionally, we have incorporated an improved attention mechanism into the backbone network, concatenated with I_CBAM, which enhances ResNeXt's ability to extract potentially critical features from fuzzy ship images. The experimental results, based on DataCastle's public dataset, indicate that the model achieves a classification accuracy of 92.86%, precision of 87.68%, and F1-Score of 81.67% when using only 5 images with 50 labelled images in each of the ten different ship images.