## Digital Image Analysis: Image Quality Metrics

• The image quality can be assessed either using subjective quality methods or objective quality methods.
• The subjective quality methods are based on the visual appearance of the image as per human perception.
• The objective quality methods are based on the availability of a reference image, application scope, and model of a Human Visual System. And based on these factors, these methods generate a numeric score that decides the image quality based on that numeric score.
The objective quality metrics are used in image processing areas like image denoising, image segmentation, image fusion, etc. to assess the input image quality. These metrics are categorized into two different sections:
1. Reference Image Quality Metrics (Full reference)
2. No Reference Image Quality Metrics (No-reference)

### No Reference Image Quality Metrics (No-reference)

These types of metrics use statistical features like probabilistic description or classification of the texture of the image to assess the quality of the image. There are various metrics used under this category such as:
• Standard Deviation (SD): It is the square root of noise variance. It analyses the contrast-level of the image. It is used in various fields of image processing like image denoising, image fusion, etc. The higher SD value indicates better perceptual image quality.
Matlab provides an inbuilt function i.e. sd_val = std2(X), where sd_val represents the SD value of image X.
• Entropy (E): It is used to characterize the texture of the input image. The higher E value indicates better perceptual image quality.
Matlab provides an inbuilt function i.e. E = entropy(X), where E represents the entropy value of grayscale image X.
• Equivalent Number of Look (ENL): It is used to analyze the smoothening feature of the image. It is evaluated over the uniform regions of the image. It is defined as the ratio between the mean squared to the variance. This metric is mainly used in the reduction of speckle noise.
In Matlab, it can be calculated as
mean_val = mean2(X)
sd_val = std2(X)
ENL = ((mean_val)∧2 / sd_val)
• Ratio Image (RI): It describes the overall information present in the image with respect to uniform and non-uniform regions of the image. It is calculated by taking the ratio of the noisy and denoised image.
In Matlab, it can be calculated as
imshow (X/Y)
Here, X and Y are the noisy and denoised images respectively.
• Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE): It draws out the probabilistic analysis of local normalized luminance signals. It measures the naturalness of the image. A smaller value of BRISQUE represents better perceptual image quality.
Matlab provides an inbuilt function i.e. bri_val = brisque(X), where bri_val represents the BRISQUE value of image X.
• Naturalness Image Quality Evaluator (NIQE): It calculates the quantifiable deflection from probabilistic constancy seen in natural images. A smaller value of NIQE represents better perceptual image quality.
Matlab provides an inbuilt function i.e. niqe_val = niqe(X), where niqe_val represents the NIQE value of image X.
• Perception-based Image Quality Evaluator (PIQE): It is used to assess the over smoothening in the image during the deblurring/ denoising process.
Matlab provides an inbuilt function i.e. piqe_val = piqe(X), where piqe_val represents the PIQE value of image X.
• Coefficient of Variation (CV): It is used for the assessment of texture preservation in the non-uniform regions of the image. It is mostly used in the field of speckle noise reduction. It is defined as the ratio of standard deviation divided by a mean value and expressed as a percentage.
In Matlab, it can be calculated as
cv = std(E2(:))/mean(E2(:))∗100;
• Target-to-Clutter Ratio (TCR): It is mostly used in the field of SAR image despeckling for the assessment of despeckled SAR image. It analyses the level of preservation of radiometric features being protected in the despeckled SAR image. The below formula calculates it.
• Noise Variance (NV): It is used to measure the noise content currently in the noisy and denoised images. Based on this score, the denoising process is analyzed.
If {NV(denoised_image) < NV(noisy_image)}, then the denoising process is properly performed.
In Matlab, it can be calculated as
NV = sqrt (std2(E2))
• Average Gradient (AG): It is used to evaluate the clarity of the image. It is mostly used in the area of image fusion. It checks the texture and contrast features being modified during the fusion process. The higher AG value indicates better perceptual image quality.
In Matlab, it can be calculated as
AG = sum(sum(Gmag))./(sqrt(2)*(size(X,1)-1)*(size(X,2)-1));
• Petrovics metric (PM): It is used to evaluate the amount of fine details (information like edges and structure of objects) being transferred from source images to the fused image. It is mostly used in the area of image fusion. The higher PM value indicates better perceptual image quality. The Matlab code is available in . The below formula calculates it.
• Mutual information (MI): It measures the fine details being transferred from source images to the fused image after the fusion process. It is mostly used in the area of image fusion. The higher MI value indicates better perceptual image quality.
• Average Pixel Intensity (API) or mean: It measures the overall information (foreground and background) of the image. The higher API value indicates better perceptual image quality.

#### References

For more image processing topics refer: https://www.gofastresearch.com/
Digital Image Analysis: No Reference Image Quality Metrics (No-reference) Reviewed by Dr. Prabhishek Singh on April 30, 2020 Rating: 5