Explanation of configurations in info.md - WolframRhodium/Super-Resolution-Zoo GitHub Wiki

All models have been adjusted such that they all accept floating point input images in shape "NCHW" (batch, channels, height, width) where the value of channels is 1 or 3, and range [0, 1].

Parameters

  1. up_scale: (uint) This parameter indicates the up-scaling factor of the whole processing pipeline. For super-resolution models, the corresponding values are usually 2, 3, 4, 8, while for other models that deal with other tasks like denoising, compression artifacts reduction or dehazing, the corresponding values are usually 1.

  2. is_rgb_model: (bool) This parameter indicates whether the corresponding models operate on RGB images (channels=3) directly. If not, the models operate on gray-scale images (channels=1), which are usually the luminance images converted from RGB images using MATLAB's rgb2ycbcr. Note however that this might not be the case for several models since they use a slightly different RGB-to-luminance formula (sorry I forgot which models met this condition). Default: True

  3. pad, crop: (uint or None) These parameters indicate spatial padding (before processing) or cropping (after processing) that are required for the corresponding processing pipeline. Default: 0

  4. pre_upscale: (bool) This parameter indicates whether input images are required to be pre-upscaled before feeding to the network. The up-scaling factor is specified by up_scale and the up-scaling algorithm is specified by resample_kernel. Models that requires such pre-upscaling includes VDSR, RED30(super_resolution), MemNet. Default: False

  5. merge_source: (bool) This parameter indicates whether the input image should be merged to model's output to obtain the final output image. If the resolution of input image is different to the output, it should be upscaled. Default: False

  6. upscale_uv: (bool) For models with is_rgb_model=False, this value specifies whether the UV components would be upscaled by the model. It has no effects on RGB models. Default: False

  7. resample_kernel: (string) This parameter indicates the resampling/rescaling/resizing algorithm used for pre-processing/post-processing. "Catmull-Rom" means imresize('method', 'bicubic') in MATLAB. Default: "Catmull-Rom"

Here are two useful posts by @xinntao about convensions in data-processing: Color conversion in SR, MATLAB bicubic imresize.

Please feel free to report any issues.