<p>In recent years, significant progress has been made in semi-supervised learning methods that leverage pseudo-labels and consistency regularization. However, two key issues with existing approaches have been identified. Firstly, these methods often focus on maintaining inter-class consistency between strong and weak augmentations, but they tend to overlook intra-class differences. This oversight results in a waste of valuable intra-class information. Secondly, the selection of a fixed high threshold for pseudo-label confidence restricts the quantity and utilization of pseudo-labels. On the contrary, using an initial low threshold introduces a large number of erroneous pseudo-labels, resulting in a degradation of the model’s performance. To address these issues, we propose a novel semi-supervised learning framework that combines contrastive learning and feature discrepancy loss. Our approach introduces a new loss function that facilitates intra-class discrimination by emphasizing inter-class differences. Additionally, we tackle the threshold problem in pseudo-label consistency loss by introducing dynamic weighting coefficients. These coefficients help balance the impact of both the quantity and quality of pseudo-labels on the model. Our experimental results demonstrate that our method effectively harnesses the feature differences among samples with different perturbations. This enhancement boosts the model’s feature generation capability while mitigating the negative effects of erroneous pseudo-labels on model’s performance. Overall, our proposed framework provides a more comprehensive and effective solution to semi-supervised learning in classification applications by addressing the issues of intra-class differences and the selection of pseudo-label thresholds.</p>