Improving Image Restoration by Revisiting Global Information Aggregation

Global operations, such as global average pooling, are widely used in top-performance image restorers. They aggregate information from input features along entire spatial dimensions but behave differently during training and inference restoration tasks: they based on different regions, namely the cropped patches (from images) full-resolution images. This paper revisits aggregation finds that image-based have a distribution than patch-based training. train-test inconsistency negatively impacts performance of models, which is severely overlooked by previous works. To reduce improve test-time performance, we propose simple method called Test-time Local Converter (TLC). Our TLC converts operations to local ones only so within regions rather large The proposed can be applied various modules (e.g., normalization, channel attention) with negligible costs. Without need for any fine-tuning, improves state-of-the-art results several tasks, including single-image motion deblurring, video defocus denoising. In particular, TLC, our Restormer-Local result single deblurring 32.92 dB 33.57 GoPro dataset. code available at https://github.com/megvii-research/tlc .