feat: 🚩 add entropy for NR IQA

ResultsCalibra/calibration_summary.csv

@@ -7,6 +7,8 @@ cw_ssim,0.2763,0.9996,1.0,0.9068,0.8658
 cw_ssim(ours),0.2763,0.9996,1.0,0.9068,0.8658
 dists,0.4742,0.1424,0.0683,0.0287,0.3123
 dists(ours),0.4742,0.1424,0.0682,0.0287,0.3123
+entropy,6.9511,6.9661,7.5309,7.5566,5.7629
+entropy(ours),6.9511,6.9661,7.5309,7.5565,5.7629
 fsim,0.689,0.9702,0.9927,0.9575,0.822
 fsim(ours),0.6891,0.9702,0.9927,0.9575,0.822
 gmsd,0.2203,0.0005,0.0004,0.1346,0.205
@@ -34,7 +36,7 @@ nlpd,0.5616,0.0195,0.0159,0.3028,0.4326
 nlpd(ours),0.5616,0.0139,0.011,0.3033,0.4335
 nrqm,1.3894,8.9394,8.9735,6.829,6.312
 nrqm(ours),1.3932,8.9419,8.9721,6.8309,6.3031
-paq2piq,44.1340,73.6015,74.3297,76.8748,70.9153
+paq2piq,44.134,73.6015,74.3297,76.8748,70.9153
 paq2piq(ours),44.1341,73.6015,74.3297,76.8748,70.9153
 pi,11.9235,3.072,2.618,2.8074,6.7713
 pi(ours),11.9286,3.073,2.6357,2.7979,6.9546

ResultsCalibra/results_official.csv

@@ -22,3 +22,4 @@ musiq-koniq,12.494,75.332,73.429,75.188,36.938
 musiq-paq2piq,46.035,72.660,73.625,74.361,69.006
 musiq-spaq,17.685,70.492,78.740,79.015,49.105
 paq2piq,44.1340,73.6015,74.3297,76.8748,70.9153
+entropy,6.9511,6.9661,7.5309,7.5566,5.7629

pyiqa/archs/entropy_arch.py

@@ -0,0 +1,58 @@
+r"""Peak signal-to-noise ratio (PSNR) Metric
+
+Created by: https://github.com/photosynthesis-team/piq
+
+Modified by: Jiadi Mo (https://github.com/JiadiMo)
+
+Refer to:
+    Wikipedia from https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio
+    QIQA from https://github.com/francois-rozet/piqa/blob/master/piqa/psnr.py
+
+"""
+
+import torch
+import torch.nn as nn
+
+from pyiqa.utils.registry import ARCH_REGISTRY
+from pyiqa.utils.color_util import to_y_channel
+
+
+def entropy(x, data_range=255., eps=1e-8, color_space='yiq'):
+    r"""Compute entropy of a gray scale image.
+    Args:
+        x: An input tensor. Shape :math:`(N, C, H, W)`.
+    Returns:
+        Entropy of the image.
+    """
+
+    if (x.shape[1] == 3):
+        # Convert RGB image to gray scale and use Y-channel
+        x = to_y_channel(x, data_range, color_space)
+
+    # Compute histogram
+    hist = nn.functional.one_hot(x.long(), num_classes=int(data_range + 1)).sum(dim=[1, 2, 3])
+    hist = hist / hist.sum(dim=1, keepdim=True)
+
+    # Compute entropy
+    score = -torch.sum(hist * torch.log2(hist + eps), dim=1)
+
+    return score
+
+
+@ARCH_REGISTRY.register()
+class Entropy(nn.Module):
+    r"""
+    Args:
+        x (torch.Tensor): image tensor with shape (B, _, H, W), range [0, 1]
+    Return:
+        score (torch.Tensor): (B, 1)
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__()
+        self.kwargs = kwargs
+
+    def forward(self, x):
+
+        score = entropy(x, **self.kwargs)
+        return score

pyiqa/default_model_configs.py

@@ -344,5 +344,11 @@
             'type': 'CLIPScore',
         },
         'metric_mode': 'NR', # Caption image similarity
+    },
+    'entropy': {
+        'metric_opts': {
+            'type': 'Entropy',
+        },
+        'metric_mode': 'NR',
     }
 })