pyiqa.archs.nrqm_arch

NRQM Metric, proposed in

Chao Ma, Chih-Yuan Yang, Xiaokang Yang, Ming-Hsuan Yang “Learning a No-Reference Quality Metric for Single-Image Super-Resolution” Computer Vision and Image Understanding (CVIU), 2017

Matlab reference: https://github.com/chaoma99/sr-metric This PyTorch implementation by: Chaofeng Chen (https://github.com/chaofengc)

Module Contents

pyiqa.archs.nrqm_arch.default_model_urls

pyiqa.archs.nrqm_arch.get_gauss_pyramid(x: torch.Tensor, scale: int = 2)

Get gaussian pyramid images with gaussian kernel.

pyiqa.archs.nrqm_arch.get_var_gen_gauss(x, eps=1e-07)

Get mean and variance of input local patch.

pyiqa.archs.nrqm_arch.gamma_gen_gauss(x: torch.Tensor, block_seg=10000.0)

General gaussian distribution estimation.

Parameters:

block_seg – maximum number of blocks in parallel to avoid OOM

pyiqa.archs.nrqm_arch.gamma_dct(dct_img_block: torch.Tensor)

Generalized gaussian distribution features

pyiqa.archs.nrqm_arch.coeff_var_dct(dct_img_block: torch.Tensor)

Gaussian var, mean features

pyiqa.archs.nrqm_arch.oriented_dct_rho(dct_img_block: torch.Tensor)

Oriented frequency features

pyiqa.archs.nrqm_arch.block_dct(img: torch.Tensor)

Get local frequency features

pyiqa.archs.nrqm_arch.norm_sender_normalized(pyr, num_scale=2, num_bands=6, blksz=3, eps=1e-12)

Normalize pyramid with local spatial neighbor and band neighbor

pyiqa.archs.nrqm_arch.global_gsm(img: torch.Tensor)

Global feature from gassian scale mixture model

pyiqa.archs.nrqm_arch.tree_regression(feat, ldau, rdau, threshold_value, pred_value, best_attri)

Simple decision tree regression.

pyiqa.archs.nrqm_arch.random_forest_regression(feat, ldau, rdau, threshold_value, pred_value, best_attri)

Simple random forest regression.

Note: currently, this is non-differentiable and only support CPU.

pyiqa.archs.nrqm_arch.nrqm(img: torch.Tensor, linear_param, rf_param) → torch.Tensor

Calculate NRQM :param img: Input image. :type img: Tensor :param linear_param: (4, 1) linear regression params :type linear_param: np.array :param rf_param: params of 3 random forest for 3 kinds of features

pyiqa.archs.nrqm_arch.calculate_nrqm(img: torch.Tensor, crop_border: int = 0, test_y_channel: bool = True, color_space: str = 'yiq', linear_param: torch.Tensor = None, rf_params_list: list = None, **kwargs) → torch.Tensor

Calculate NRQM :param img: Input image whose quality needs to be computed. :type img: Tensor :param crop_border: Cropped pixels in each edge of an image. These

pixels are not involved in the metric calculation.

Parameters:

• test_y_channel (Bool) – Whether converted to ‘y’ (of MATLAB YCbCr) or ‘gray’.

• pretrained_model_path (String) – The pretrained model path.

Returns:

NIQE result.

Return type:

Tensor

class pyiqa.archs.nrqm_arch.NRQM(test_y_channel: bool = True, color_space: str = 'yiq', crop_border: int = 0, pretrained_model_path: str = None)

Bases: torch.nn.Module

NRQM metric

Ma, Chao, Chih-Yuan Yang, Xiaokang Yang, and Ming-Hsuan Yang. “Learning a no-reference quality metric for single-image super-resolution.” Computer Vision and Image Understanding 158 (2017): 1-16.

Parameters:

• channels (-) – Number of processed channel.

• test_y_channel (-) – whether to use y channel on ycbcr.

• crop_border (-) – Cropped pixels in each edge of an image. These pixels are not involved in the metric calculation.

• pretrained_model_path (-) – The pretrained model path.

forward(X: torch.Tensor) → torch.Tensor

Computation of NRQM metric. :param X: An input tensor. Shape \((N, C, H, W)\).

Returns:

Value of nrqm metric.

class pyiqa.archs.nrqm_arch.PI(crop_border=4, color_space='ycbcr')

Bases: torch.nn.Module

Perceptual Index (PI), introduced by

Blau, Yochai, Roey Mechrez, Radu Timofte, Tomer Michaeli, and Lihi Zelnik-Manor. “The 2018 pirm challenge on perceptual image super-resolution.” In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pp. 0-0. 2018. Ref url: https://github.com/roimehrez/PIRM2018

It is a combination of NIQE and NRQM: 1/2 * ((10 - NRQM) + NIQE)

Parameters:

• color_space (-) – color space of y channel, default ycbcr.

• crop_border (-) – Cropped pixels in each edge of an image, default 4.

forward(X: torch.Tensor) → torch.Tensor

Computation of PI metric. :param X: An input tensor. Shape \((N, C, H, W)\).

Returns:

Value of PI metric.