Reproduce “Learning from History: Task-agnostic Model Contrastive Learning for Image Restoration” on OpenResearch

About this paper

Gang Wu, Junjun Jiang, Kui Jiang, Xianming Liu

Researchers at the Harbin Institute of Technology introduce a task-agnostic model contrastive learning paradigm for image restoration, leveraging 'latency models' (earlier versions of the training model) to dynamically generate negative samples. This approach consistently improves the performance of existing CNN and Transformer-based image restoration models across various tasks, showing gains up to 3.41 dB in PSNR for image dehazing.

Key insights

• Leveraging a model's own evolutionary history through latency models provides an adaptive and task-agnostic source of challenging negative samples for contrastive learning in image restoration.
• Shifting the focus from sample-level to model-level contrastive learning effectively overcomes the limitations of task-oriented biases, allowing for a more generalizable enhancement strategy for diverse image restoration tasks and model architectures.
• The Self-Prior Guided Negative Loss acts as an implicit regularization mechanism, pushing the current model's output away from its past, suboptimal reconstructions, which consistently improves final image quality.

Summary by alphaXiv.