Enhancing Test-Time Training for Corruption Robustness in Image Classification

초록

Deep neural networks for image classification often suffer significant performance degradation when faced with common image corruptions, limiting their real-world applicability. Test Time Training (TTT) is a powerful adaptation paradigm that addresses this challenge, but its effectiveness can be constrained by the underlying model architecture and adaptation strategy. In this paper, we propose a novel framework that significantly enhances corruption robustness by synergistically combining three key strategies that span the training and testing phases. First, we adapt the original TTT architecture for image classification using a bidirectional, gated design to effectively process non-causal spatial features. Second, during training, we employ Mixup augmentation to learn more generalizable and robust feature representations, which serve as a superior foundation for adaptation. Third, at test time, we introduce an entropy-based sample ordering strategy that creates an effective "hard-to-easy"curriculum, forcing the model to adapt to the most uncertain samples first. Our final model achieves an average accuracy of 92.86% on the CIFAR10-C benchmark, demonstrating a significant improvement over a strong baseline. Our results show that this multi-faceted approach is a crucial step towards building more reliable image classification models.

키워드