LION.metrics.haarpsi

Classes

HAARPsi([preprocess_with_subsampling, C, a])

HaarPSI: the Haar wavelet-based perceptual similarity index, for (potentially) batched pytorch tensors, used like LION's PSNR and SSIM.

---

class LION.metrics.haarpsi.HAARPsi(preprocess_with_subsampling=True, C=30.0, a=4.2)

Bases: Module

HaarPSI: the Haar wavelet-based perceptual similarity index, for (potentially) batched pytorch tensors, used like LION’s PSNR and SSIM.

Reisenhofer, Bosse, Kutyniok and Wiegand, “A Haar wavelet-based perceptual similarity index for image quality assessment”, Signal Processing: Image Communication 61 (2018). PyTorch implementation by Sören Dittmer and Clemens Karner (through project FWF T1307-N/Anna Breger), from the TensorFlow implementation by David Neumann and the MATLAB one by Rafael Reisenhofer. It matches piq.haarpsi for images in [0, data_range].

preprocess_with_subsampling: subsample the images by 2 first, as in the paper (True); C, a: the algorithm’s constants, found experimentally by HaarPSI’s authors (30 and 4.2, for images in [0, 255]).

Initialize internal Module state, shared by both nn.Module and ScriptModule.

higher_is_better = True

forward(x, target, reduce=None, batched=True, data_range=None)

The HaarPSI of x against target (the same shape), in [0, 1], 1 for identical images: one per sample, or their mean if reduce is “mean”.

batched: x is (B, C, H, W) or (B, H, W), else (C, H, W) or (H, W); C is 1 (grayscale) or 3 (RGB). data_range: images are in [0, data_range] (e.g. 1), as in piq.haarpsi; None: each pair is mapped by its target’s minimum and range to a target in [0, 1], like the range PSNR and SSIM use. Images may have values outside the range (e.g. noisy reconstructions).

maps(x, target, batched=True, data_range=None)

HaarPSI with its local similarity maps and weights.

Same arguments as forward. Returns (score, local_similarities, weights): score has shape (B,), the others (B, n_channels, H, W), with H and W halved if subsampling and n_channels 2 for grayscale or 3 for colour images (always batched).

logit(value)

The inverse of the logistic function that HaarPSI’s similarities go through, scaled by 1 / a.