Learning Cascaded Convolutional Networks for Blind Single Image Super-Resolution
Thispaperstudiestheblindsuper-resolutionofreallow-qualityandlow-resolution (LR) images. Existing convolutional network (CNN) based approaches usually learn a single image super-resolution (SISR) model for a specific downsampler (e.g., bicubic downsampling, blurring followed by downsampling). The learned model, however, is tailored to the specific downsampler and fails to super-resolve real LR images which are degraded in more sophisticated and diverse manners. Moreover, the ground-truth high-resolution (HR) of real LR images are generally unavailable. Instead of learning from unpaired real LR-HR images or a specific downsampler, this paper learns blind SR network from a realistic, parametric degradation model by considering blurring, noise, downsampling, and even JPEG compression. In contrast to direct blind reconstruction of HR image, the proposed model adopts a cascaded architecture for noise estimation, blurring estimation, and non-blind SR, which can be jointly end-to-end learned from training data and benefit generalization ability. By taking the bicubicly upscaled LR image as input to non-blind SR, the proposed method can present a single unified model for blind SR with any upscaling factors and varying degradation parameters. Experimental results show that the proposed method performs favorably on synthetic and real LR images.
- Python 3.5
- PyTorch >= 1.2.0
- scipy 1.0.0
Data Set: DIV2K 800 training images.
- Pretraining model can be found here.
- Test commond
--model CBSR --save CBSR --scale 4 --n_feats 64 --save_results --print_model --n_GPUs 1 --testpath your_testset_path --ext_tt .png --n_resblocks 2 --resume -1 --pre_train your_model_path --test_only
- model.common.cov2pca ----- convert covariance matrix to kernels and reducing by PCA
- data.common.im_process ----- image degradation processing
Please send email to [email protected]