StructuralSimilarity¶

class mmeval.metrics.StructuralSimilarity(crop_border: int = 0, input_order: str = 'CHW', convert_to: Optional[str] = None, channel_order: str = 'rgb', **kwargs)[source]¶

Calculate StructuralSimilarity (structural similarity).

Ref: Image quality assessment: From error visibility to structural similarity

The results are the same as that of the official released MATLAB code in https://ece.uwaterloo.ca/~z70wang/research/ssim/.

For three-channel images, StructuralSimilarity is calculated for each channel and then averaged.

Parameters

crop_border (int) – Cropped pixels in each edges of an image. These pixels are not involved in the PeakSignalNoiseRatio calculation. Defaults to 0.
input_order (str) – Whether the input order is ‘HWC’ or ‘CHW’. Defaults to ‘HWC’.
convert_to (str, optional) – Whether to convert the images to other color models. If None, the images are not altered. When computing for ‘Y’, the images are assumed to be in BGR order. Options are ‘Y’ and None. Defaults to None.
channel_order (str) – The channel order of image. Choices are ‘rgb’ and ‘bgr’. Defaults to ‘rgb’.
**kwargs – Keyword parameters passed to BaseMetric.

Examples

>>> from mmeval import StructuralSimilarity as SSIM
>>> import numpy as np
>>>
>>> ssim = SSIM(input_order='CHW', convert_to='Y', channel_order='rgb')
>>> gts = np.random.randint(0, 255, size=(3, 32, 32))
>>> preds = np.random.randint(0, 255, size=(3, 32, 32))
>>> ssim(preds, gts)  
{'ssim': ...}

Calculate StructuralSimilarity between 2 single channel images:

>>> img1 = np.ones((32, 32)) * 2
>>> img2 = np.ones((32, 32))
>>> SSIM.compute_ssim(img1, img2)
0.913062377743969

add(predictions: Sequence[numpy.ndarray], groundtruths: Sequence[numpy.ndarray], channel_order: Optional[str] = None) → None[source]¶

Add the StructuralSimilarity score of the batch to self._results.

For three-channel images, StructuralSimilarity is calculated for each channel and then averaged.

Parameters

predictions (Sequence[np.ndarray]) – Predictions of the model.
groundtruths (Sequence[np.ndarray]) – The ground truth images.
channel_order (Optional[str]) – The channel order of the input samples. If not passed, will set as self.channel_order. Defaults to None.

compute_metric(results: List[numpy.float64]) → Dict[str, float][source]¶

Compute the StructuralSimilarity metric.

This method would be invoked in BaseMetric.compute after distributed synchronization.

Parameters: results (List[np.float64]) – A list that consisting the StructuralSimilarity scores. This list has already been synced across all ranks.
Returns: The computed StructuralSimilarity metric.
Return type: Dict[str, float]

static compute_ssim(img1: numpy.ndarray, img2: numpy.ndarray) → numpy.float64[source]¶

Calculate StructuralSimilarity (structural similarity) between two single channel image.

Ref: Image quality assessment: From error visibility to structural similarity

The results are the same as that of the official released MATLAB code in https://ece.uwaterloo.ca/~z70wang/research/ssim/.

Parameters

img1 (np.ndarray) – Single channels Images with range [0, 255].
img2 (np.ndarray) – Single channels Images with range [0, 255].

Returns

StructuralSimilarity result.

Return type

np.float64