Image Quality Metrics

[Margurite Vizarro]

Imagequality should not be mistaken with image fidelity. Image fidelity refers to the ability of a process to render a given copy in a perceptually similar way to the original (without distortion or information loss), i.e., through a digitization or conversion process from analog media to digital image.

The image formation process is affected by several distortions between the moment in which the signals travel through to and reach the capture surface, and the device or mean in which signals are displayed. Although optical aberrations can cause great distortions in image quality, they are not part of the field of Image Quality Assessment. Optical aberrations caused by lenses belong to the optics area and not to the signal processing areas.

In an ideal model, there's no quality loss between the emission of the signal and the surface in which the signal is being captured on. For example, a digital image is formed by electromagnetic radiation or other waves as they pass through or reflect off objects. That information is then captured and converted into digital signals by an image sensor. The sensor, however, has nonidealities that limit its performance.

Image quality can be assessed using objective or subjective methods. In the objective method, image quality assessments are performed by different algorithms that analyze the distortions and degradations introduced in an image. Subjective image quality assessments are a method based on the way in which humans experience or perceive image quality. Objective and subjective methods of quality assessment don't necessarily correlate with each other. An algorithm might have a similar value for an image and its altered or degraded versions, while a subjective method might perceive a stark contrast in quality for the same image and its versions.

Subjective methods for image quality assessment belong to the larger area of psychophysics research, a field that studies the relationship between physical stimulus and human perceptions. A subjective IQA method will typically consist on applying mean opinion score techniques, where a number of viewers rate their opinions based on their perceptions of image quality. These opinions are afterwards mapped onto numerical values.

Since visual perception can be affected by environmental and viewing conditions, the International Telecommunication Union produced a set of recommendations for standardized testing methods for subjective image quality assessment.[4]

Image quality metrics can also be classified in terms of measuring only one specific type of degradation (e.g., blurring, blocking, or ringing), or taking into account all possible signal distortions, that is, multiple kinds of artifacts.[7]

Efforts have been made to create objective measures of quality. For many applications, a valuable quality metric correlates well with the subjective perception of quality by a human observer. Quality metrics can also track unperceived errors as they propagate through an image processing pipeline, and can be used to compare image processing algorithms.

If an image without distortion is available, you can use it as a reference to measure the quality of other images. For example, when evaluating the quality of compressed images, an uncompressed version of the image provides a useful reference. In these cases, you can use full-reference quality metrics to directly compare the target image and the reference image.

Structural similarity (SSIM) index. The SSIM metric combines local image structure, luminance, and contrast into a single local quality score. In this metric, structures are patterns of pixel intensities, especially among neighboring pixels, after normalizing for luminance and contrast. Because the human visual system is good at perceiving structure, the SSIM quality metric agrees more closely with the subjective quality score.

The BRISQUE and the NIQE algorithms calculate the quality score of an image with computational efficiency after the model is trained. PIQE is less computationally efficient, but it provides local measures of quality in addition to a global quality score. All no-reference quality metrics usually outperform full-reference metrics in terms of agreement with a subjective human quality score.

Quality metrics provide an objective score of image quality. Full reference algorithms compare the input image against a pristine reference image with no distortion. No-reference algorithms compare statistical features of the input image against a set of features derived from an image database.

Standardized test charts contain visual features, such as slanted edges, gray patches, and color patches. These features enable the measurement of corresponding image quality characteristics, such as sharpness and color accuracy.

Image Quality Assessment (IQA) is considered as a characteristic property of an image. Degradation of perceived images is measured by image quality assessment. Usually, degradation is calculated compared to an ideal image known as reference image.

Quality of image can be described technically as well as objectively to indicate the deviation from the ideal or reference model. It also relates to the subjective perception or prediction of an image [1] , such as an image of a human look.

There are several techniques and metrics available to be used for objective image quality assessment. These techniques are grouped into two categories based on the availability of a reference image [3] . The categories are as follows:

1) Full-Reference (FR) approaches: The FR approaches focus on the assessment of the quality of a test image in comparison with a reference image. This reference image is considered as the perfect quality image that means the ground truth. For example, an original image is compared to the JPEG-compressed image [3] [4] .

The mean squared error (MSE) is the most widely used and also the simplest full reference metric which is calculated by the squared intensity differences of distorted and reference image pixels and averaging them with the peak signal-to-noise ratio (PSNR) of the related quantity [5] .

Image quality assessment metrics such as MSE, PSNR are mostly applicable as they are simple to calculate, clear in physical meanings, and also convenient to implement mathematically in the optimization context. But they are sometimes very mismatched to perceive visual quality and also are not normalized in representation. With this view, researchers have taken-into account, two normalized reference methods to give structural and feature similarities. Structured similarity indexing method (SSIM) gives normalized mean value of structural similarity between the two images and feature similarity indexing method (FSIM) gives normalized mean value of feature similarity between the two images. All these are full-reference image quality measurement metrics.

There are so many image quality techniques largely used to evaluate and assess the quality of images such as MSE (Mean Square Error), UIQI (Universal Image Quality Index), PSNR (Peak Signal to Noise Ratio), SSIM (Structured Similarity Index Method), HVS (Human Vision System), FSIM (Feature Similarity Index Method), etc. In this paper, we have worked on SSIM, FSIM, MSE and PSNR methods to find their suitability.

It is the second moment of the error. The variance of the estimator and its bias are both incorporated with mean squared error. The MSE is the variance of the estimator in case of unbiased estimator. It has the same units of measurement as the square of the quantity being calculated like as variance. The MSE introduces the Root-Mean-Square Error (RMSE) or Root-Mean-Square Deviation (RMSD) and often referred to as standard deviation of the variance.

The MSE can also be said the Mean Squared Deviation (MSD) of an estimator. Estimator is referred as the procedure for measuring an unobserved quantity of image. The MSE or MSD measures the average of the square of the errors. The error is the difference between the estimator and estimated outcome. It is a function of risk, considering the expected value of the squared error loss or quadratic loss.

Root Mean square Error is another type of error measuring technique used very commonly to measure the differences between the predicted value by an estimator and the actual value. It evaluates the error magnitude. It is a perfect measure of accuracy which is used to perform the differences of forecasting errors from the different estimators for a definite variable [7] .

PSNR is used to calculate the ratio between the maximum possible signal power and the power of the distorting noise which affects the quality of its representation. This ratio between two images is computed in decibel form. The PSNR is usually calculated as the logarithm term of decibel scale because of the signals having a very wide dynamic range. This dynamic range varies between the largest and the smallest possible values which are changeable by their quality.

The Peak signal-to-noise ratio is the most commonly used quality assessment technique to measure the quality of reconstruction of lossy image compression codecs. The signal is considered as the original data and the noise is the error yielded by the compression or distortion. The PSNR is the approximate estimation to human perception of reconstruction quality compared to the compression codecs.

In image and video compression quality degradation, the PSNR value varies from 30 to 50 dB for 8-bit data representation and from 60 to 80 dB for 16-bit data. In wireless transmission, accepted range of quality loss is approximately 20 - 25 dB [8] .

Here, peakval (Peak Value) is the maximal in the image data. If it is an 8-bit unsigned integer data type, the peakval is 255 [8] . From Equation (3), we can see that it is a representation of absolute error in dB.

Structural Similarity Index Method is a perception based model. In this method, image degradation is considered as the change of perception in structural information. It also collaborates some other important perception based fact such as luminance masking, contrast masking, etc. The term structural information emphasizes about the strongly inter-dependant pixels or spatially closed pixels. These strongly inter-dependant pixels refer some more important information about the visual objects in image domain. Luminace masking is a term where the distortion part of an image is less visible in the edges of an image. On the other hand contrast masking is a term where distortions are also less visible in the texture of an image. SSIM estimates the perceived quality of images and videos. It measures the similarity between two images: the original and the recovered.

3a8082e126