Fine-grained adaptive contrastive learning for unsupervised feature extraction

Self-supervised contrastive learning (SCL) has emerged as a crucial technique for unsupervised fine-grained feature extraction in the deep learning field. However, model collapse is a prevalent issue that hinders the potential of this approach, particularly in fine-grained tasks. To address the performance degradation in SCL, we propose a novel framework called Fine-grained Adaptive Contrastive Learning (FACL), which consists of three carefully designed modules. The first module is the adaptive center-point regression, which empowers the segment anything model to extract regions of interest adaptively. The second module, Diffusion Gaussian Noise, creates proxy distributions for the original samples, thereby augmenting the data with increased diversity. Finally, the lightweight multi-scale instant extraction module improves the efficiency and accuracy of feature extraction by integrating multi-scale sensory fields. Extensive experiments on real-world datasets demonstrate the effectiveness of the FACL framework in various tasks, including biometric authentication and anomaly detection. FACL outperforms existing SCL methods, reducing the equal error rate up to 54.8% and 64.1% compared to SimCLR and BYOL, respectively.