Robust Contrastive Learning: PSSI - Image Entropy Positive Samples with Semantic Invariance

Data augmentation is a popular technique used to increase the amount of training data and improve model generalization in the contrastive learning framework. However, traditional enhancement methods often struggle to balance the preservation of semantic information and the enhancement of generalization ability. In this paper, we propose a lightweight data augmentation method called Positive Samples with Semantic Invariance (PSSI) that aims to preserve image semantics and model generalization to the greatest extent possible. PSSI uses an innovative edge blurring technique to emphasize the model’s learning of high-level semantic information while maintaining image entropy. We show that PSSI outperforms other methods without increasing model complexity and is more robust in sparse sample scenarios. Specifically, our experiments on popular image classification datasets demonstrate that PSSI improves model performance by 1.72% with no missing samples and 4.05% with missing samples. Our results highlight the effectiveness of PSSI in balancing the preservation of semantic information and the enhancement of generalization ability, which can help improve model performance in various applications.