Read the full white paper (PDF)

White paper by Norman Koren (Imatest LLC) — May 29, 2026

Overview

This white paper is a concise, practical guide to camera image quality testing — with an emphasis on machine vision — built around metrics derived from information theory, which predict real-world camera performance better than traditional metrics such as sharpness (MTF) or noise alone.

It introduces Information Capacity (C) as the key performance metric, derived from sharpness, noise, and signal amplitude, and C4, the information capacity measured directly from 4:1-contrast slanted edges (the maximum information per pixel for an object with 4:1 contrast). From C4 and object size it develops the ideal-observer signal-to-noise ratio (SNRi) and Object Detection Error Probability (ODEP), and presents the new Information-based Dynamic Range (InfoDR) test chart and its plot of C4 versus illumination for characterizing camera performance across a wide range of light levels, including low light and dynamic range.

Along the way it reviews the underlying imaging concepts (raw conversion and demosaicing, gamma and tonal response, color spaces, MTF/SFR and slanted-edge charts, noise, and lighting), walks through spatial (slanted-edge) and tonal camera characterization with the relevant Imatest charts — including the 36-patch Dynamic Range (DR36), Contrast Resolution, and InfoDR charts — and covers the Photon Transfer Curve and the Simatest camera/ISP simulator.

It closes with color, uniformity (light falloff) and defect (blemish) measurements, and a discussion of problematic images and misleading results, including obsolete test charts, oversharpening, bilateral filtering and texture loss, and tone mapping.

See also

• Image Information Metrics
• Full PDF
• Imatest InfoDR Chart
• Using InfoDR: Information-based Dynamic Range, Part 1
• Using InfoDR, Part 2: Chart analysis
• InfoDR (Information-based Dynamic Range) Part 3: Results
• Image information metrics from Slanted edges: Instructions

The post Image Quality Testing Based on Information Metrics appeared first on Imatest.

Read the full paper (PDF)

Authors: Christian Taylor and Amelia Limbocker (Imatest LLC)

Presented March 4, 2026 at the IS&T Electronic Imaging Symposium 2026, in the AVM (Autonomous Vehicles and Machines) session.

Abstract

Near-infrared (NIR) imaging is now prevalent in machine vision, automotive, and biomedical applications, but most step-chart definitions were created for visible imaging. Popular standards assume visible-band weighting and do not account for NIR-sensitive systems. This leads to charts with target densities unoptimized for NIR applications.

We present a methodology for designing NIR test charts whose optical densities (ODs) align with the effective bandpass of a specific camera. First, we model the ISO 5 visual density calculation and the ISO 14524 OECF method for determining density targets for a test chart from the spectra of an inkjet print. That method is modified to accept new bandpass weights to accommodate sensitivities not accounted for by ISO 5-3. Three new weights are applied, and reflectance factor density targets are calculated. The results show that target densities shift with alternative spectral weightings, motivating a different NIR chart design.

See also

• Imatest at Electronic Imaging 2026
• Full paper (PDF)

The post A Method for Calculating NIR Bandpass-Adjusted Optical Densities for Better Matching Common Standard Test Chart Specifications appeared first on Imatest.

Read the full paper (PDF)

Author: Norman L. Koren (Imatest LLC)

Presented March 4, 2026 at the IS&T Electronic Imaging Symposium 2026, in the AVM (Autonomous Vehicles and Machines) session.

Abstract

We present a new method for measuring a camera’s Dynamic Range (DR) and low light performance, both of which are derived from C4 information capacity, which is measured directly from ISO 12233-standard 4:1 contrast slanted edges.

The method uses a new test chart that consists of groups of squares in a compact arrangement, where each square differs in transmittance or reflectance from its neighbors by a factor of 4, so that all edges between adjacent squares in a group have a 4:1 contrast ratio (a density step of 0.602).

The major advantage of C4 is that it completely characterizes the performance of cameras for objects with 4:1 contrast, whereas the traditional metrics — signal amplitude, sharpness, and noise, each of which contributes to information capacity — do not individually constitute complete camera performance metrics.

Because the new technique uses the difference in Digital Numbers (DNs) across an edge as the signal for calculating C4, it avoids a measurement issue with simple flat patches, where stray light can be misinterpreted as improved Signal-to-Noise Ratio (SNR), distorting the measurements. It does, however, require that the test chart be well-focused. (The old technique was tolerant of moderate misfocus.)

Finally, we examine a new plot of C4 as a function of exposure, which is a superior representation of camera performance over a wide range of illumination.

See also

• Imatest at Electronic Imaging 2026
• Full paper (PDF)
• Image Quality Testing Based on Information Metrics
• Imatest InfoDR Chart
• Using InfoDR: Information-based Dynamic Range, Part 1
• Using InfoDR, Part 2: Chart analysis
• InfoDR (Information-based Dynamic Range) Part 3: Results
• Image Information Metrics

The post Information-based Dynamic Range appeared first on Imatest.

Read the full paper (PDF)

Author: Sarah Kerr (Imatest LLC)

Presented March 4, 2026 at the IS&T Electronic Imaging Symposium 2026, in the IQSP session.

Abstract

Edge localization plays a critical role in ISO 12233 e-SFR analysis, influencing both sharpness results and downstream information capacity metrics. This paper evaluates the accuracy of the standard centroid, low-pass filter, and matched filter-based localization methods across an ensemble of simulated slanted edge ROIs. Localization errors are quantified by benchmarking each method against ground truth, and their propagation to e-SFR results and information capacity is measured. Findings show that centroid fitting introduces angular bias under noise, leading to a degraded effective response, while low-pass filtering and matched filtering both maintain robust accuracy. These results highlight an under-characterized source of error in standards-based image quality analysis and provide a foundation for improved methods. The results support a closer alignment between edge analysis, information-theoretic models, and emerging metrics such as those proposed in ISO/WD 23654 (Digital Imaging — Information Metrics).

See also

• Imatest at Electronic Imaging 2026
• Full paper (PDF)

The post From Centroid to Low-Pass Edge Fitting in ISO 12233 eSFR: Accuracy and Impact on Digital Imaging Information Metrics appeared first on Imatest.

Read the full paper (PDF)

Author: Norman L. Koren (Imatest LLC)

Presented March 3, 2026 at the IS&T Electronic Imaging Symposium 2026, in the IQSP session.

Abstract

We present an image sensor noise model that can be used in a complete image system simulation that includes image generation, lens degradations, and ISP (Image Signal Processing), and can produce classic measurements (SFR, noise, etc.) as well as new information metrics such as information capacity and SNRi.

The noise model is derived from a classic Photon Transfer Curve (PTC) obtained from one or more raw (undemosaiced) images of a high dynamic range grayscale test chart. Image sensor noise is composed of three factors: (1) dark noise, which includes electronic noise, dark current noise, DSNU fixed-pattern noise, and noise from several other sources, and is independent of signal amplitude, A; (2) photon shot noise, which varies with √A; and (3) PRNU fixed-pattern noise, which varies linearly with A.

The coefficients for the three factors are determined using a Levenberg–Marquardt optimizer that provides an extremely close fit between the measured data and the calculated PTC. The coefficients can also be derived from EMVA 1288 measurements, which are more accurate and detailed, but require the acquisition of a large number of images. We show how the model can predict performance over a wide range of conditions, and most importantly, for low light.

See also

• Imatest at Electronic Imaging 2026
• Full paper (PDF)
• Simatest Camera/ISP Simulator Overview
• Simatest Camera/ISP simulator examples
• Simatest Camera/ISP Simulator Documentation
• Camera performance simulator based on information theory

The post Image Sensor Noise Model for Image System Simulation appeared first on Imatest.

Read the full paper (PDF)

Authors: Megan Borek and Amelia Limbocker (Imatest LLC); Ellis Monk (Harvard University)

Presented March 3, 2026 at the IS&T Electronic Imaging Symposium 2026, in the IQSP session “Skin Tone Capture and Image Quality I”.

Abstract

Photographic test charts for measuring color accuracy in cameras have historically included a limited number of skin tones, typically in the form of uniform color patches. Such charts are not representative of the wide range of skin tones found in humans, and do not test the behavior of modern automatic exposure, white balance, and focus (3A) algorithms that are commonly driven by facial detection in today’s digital consumer cameras. We built upon our previous work on the development of printed skin tone charts featuring detectable faces by conducting a study with human participants whose skin tones approximately span the Monk Skin Tone Scale. Participants were photographed under a series of controlled lighting conditions, and each scene was then reproduced using a high-resolution inkjet print of the participant. Corresponding captures of the human subjects and the printed charts were quantitatively compared by calculating the CIEDE2000 color difference for regions of interest across the subject’s face in the scene. This analysis evaluates how printed skin tones behave across exposure settings and lighting conditions relative to real skin, with the goal of determining whether printed charts provide a suitable solution for repeatable, lab-based image quality testing in face-present scenes. While not intended to replace final field testing with real human subjects, results indicate that face charts printed with sufficiently wide-gamut printers can provide an effective solution for lab testing and benchmarking of color accuracy and 3A behavior in a controlled and repeatable manner.

See also

• Imatest at Electronic Imaging 2026
• Full paper (PDF)
• Boulder image quality lab addresses racial bias in consumer cameras
• Equity in Camera Technologies: How Consumer Cameras Perform Across Skin Tones
• Diverse Skin Tone Face Targets (Set of 10)-V2

The post Toward Fair and Accurate Camera Testing: Validation of Skin Tone Test Charts with Real Human Data appeared first on Imatest.

Read the full paper (open access)

Authors: Diarmaid Geever, Tim Brophy, Dara Molloy, Roshan George, Martin Glavin, Edward Jones, and Brian Deegan (University of Galway / Ryan Institute / Lero); Enda Ward (Valeo); and Norman Koren (Imatest LLC).

Published January 16, 2026 in the IEEE Open Journal of Vehicular Technology (open access). DOI: 10.1109/OJVT.2026.3655075.

Abstract

The design of automated driving systems is of growing industry and societal interest. Perception is a critical technology for these systems, which allows a vehicle to discern the surrounding environment. Perception systems in automated vehicles frequently use machine vision algorithms; however, the performance of a machine vision algorithm critically depends on the quality of the data provided. Quantifying the “quality” of image data is therefore potentially a useful tool in understanding and predicting the performance of a machine vision system. This study uses the Shannon Information Capacity, a metric based on information theory, to evaluate the impact of image quality on a perception algorithm. In this preliminary study, a set of synthetic objects are arranged to create a novel simulated test chart. The chart contains standard machine vision objects of interest (people and cars) as well as a slanted edge, which is used to calculate image quality metrics. The chart is degraded using varying levels of contrast and blur to simulate different real-world operating conditions. Object detection performance is then evaluated using a range of deep learning-based detection algorithms, with different architectures. The results indicate that Shannon Information Capacity has the potential to predict machine vision performance across multiple model architectures and object types. For example, the results for all the models show that accuracy remains relatively constant above an SIC value of 0.25 b/p. Results indicate that for YOLOv10 m SIC has mutual information value with detection accuracy of 1.66 bits while MTF50 has a score of 0.4945 bits. This study is the first to show the correlation between SIC and machine vision performance. While other metrics have been previously shown to have some correlation with machine vision, the correlation shown by SIC is much stronger. The findings presented may be of use to designers of autonomous driving systems and automotive camera manufacturers.

See also

• Image Information Metrics
• DOI: 10.1109/OJVT.2026.3655075

The post Information Capacity as a Predictor of Perception Performance appeared first on Imatest.

Norman Koren describes a camera performance simulator that includes an image sensor noise model with parameters derived from either dynamic range test chart measurements or EMVA 1288 results. The simulator includes lens degradations and the effects of Image Signal Proces­sing (ISP). Results can be analyzed for standard image quality metrics such as sharpness and noise, or new metrics, derived from information theory, that provide an improved prediction of machine vision system performance.

The image sensor noise model can be derived from a photon transfer curve (PTC) [1] obtained from a raw (undemosaiced) image of a transmissive grayscale dynamic range test chart, taking advantage of a valuable property of raw images: the mean value of a flat patch is independent of color.

Photon Transfer Curve (PTC)

The Photon Transfer Curve, shown above, contains the noise in each patch in the chart as a function of the normalized input RGB level (the normalized Digital Number, DN_n = DN/DN_max), and a gray line with a very close fit to the data, derived from an equation based on

1. the total dark noise, which consists of temporal and DSNU (Dark Signal NonUniformity, i.e., dark fixed pattern) noise,
2. the photon shot noise, which tends to dominate the response, and is proportional to the square root of DN_n, and
3. the PRNU (Photo Response NonUniformity, i.e., light fixed pattern) noise, which is proportional to DN_n).

The coefficients for the three noise factors are found with a Levenberg Marquardt optimizer that minimizes an array representing the differences between the individual patch noise minus the curve fit, divided by the individual patch noise.

EMVA 1288-based parameters can also be used to derive the noise coefficients.

1. The total dark noise coefficient is the square root of the sum of the squares of the temporal dark noise, DSNU, and dark current noise.
2. The photon shot noise coefficient is based on the gain (in DN per electron) and the maximum DN for the camera.
3. The PRNU coefficient is PRNU(%)/100.

Camera simulator and information metrics — Input to the simulator consists of ideal test chart images or acquired camera images. Ideal test chart images may be degraded by lens design programs or by blur measured from camera images. The image sensor noise model described above is applied, followed by Image signal processing (ISP).

The output of the simulator is sent to Imatest modules (Rescharts and Color/Tone) that analyze tra­di­tional image quality metrics such as sharpness (MTF), noise, SNR, color accuracy, tonal response, and dyna­mic range, as well as a new set of metrics based on information theory, originally deve­loped by Claude Shannon in 1948 [2]. Information is defined as the resolution of uncertain­ty, which can be the result from an experiment such as a coin flip or data trans­mis­sion. The key metric, (Shannon) Information Capacity, SIC, which combines SNR and sharpness (MTF), is the maximum rate that information can pass through a communi­cation channel without error. It is calculated from the Shannon-Hartley equation.

An image is such a communication channel. It can be characterized by SIC in units of bits per pixel. Because it was cumbersome to measure, it failed to gain traction in the imaging industry. In 2022 we discovered a way to conveniently measure SIC and related metrics from the widely used slanted-edge test pattern [3]. We are working with several European universities to correlate SIC with machine vision perfor­mance, and we are also working on a standard, ISO 23654. We invite participants.

Summary — We show how a subset of EMVA 1288 image sensor measurements or PTC measurements can be entered into the Simatest simulator and used to calculate camera Information Capacity (SIC) and related performance metrics, which should have a better correlation with machine vision perfor­mance than traditional metrics such as sharpness and noise or SNR.

See Also

• Simatest Overview
• [1] J. Janesick, “Photon Transfer DN → λ”, SPIE Press, 2007
• [2] C. E. Shannon, “A mathematical theory of communication”, Bell Syst. Tech. J., vol. 27, 1948
• [3] N. L. Koren, “Image Information Metrics From Slanted Edges: A Toolkit of Metrics to Aid Object Recognition, Machine Vision, and Artificial Intelligence Systems” in Electronic Imaging, 2024, pp 256-1 – 256-17, https://doi.org/10.2352/EI.2024.36.9.IQSP-256
• Image Information Metrics
• ISO 24942 (EMVA 1288)
• EMVA Forum 2025

The post Camera performance simulator based on information theory appeared first on Imatest.

Update: the contents of this post led to the publication of this paper:

D. Geever, T. Brophy, D. Molloy, E. Ward, R. George, N. Koren, M. Glavin, E. Jones, and B. Deegan,
“Information Capacity as a Predictor of Perception Performance” IEEE Open Journal of Vehicular Technology, 2024. doi: 10.1109/OJVT.2024.0627000

We are in the age of Artificial Intelligence that depends on machine vision. This technology surge has necessitated thinking about camera systems in new ways because machine vision systems based on neural networks have different requirements from camera systems intended for human perception. 

Norman Koren, Imatest’s founder, has developed a novel approach for evaluating the quality of machine vision systems based on information theory, which is widely used in electronic communications. The image signal and noise (or SNR) are entered into the Shannon-Hartley equation to calculate Information capacity, which is the maximum amount of information that can pass through a channel (i.e., the camera system) without error. Simply put, it is a measure of a camera’s “goodness.” By measuring the signal and noise from the same slanted edge region of an image, greater understanding can be gained for a camera system’s fundamental capabilities and potential performance. Traditional measurements such as SNR and MTF50 by themselves are not indicative of the fundamental image information, and hence they are insufficient to derive machine vision performance metrics such as detection confidence levels.

At Imatest, we are working with partners in industry and academia to validate the ties between a camera system’s image information metrics and its machine vision capabilities., which are based on the AI models that are trained for the system’s camera captures. We want to highlight the University of Galway (Ireland) research group led by Dr. Brian Deegan, whose lab is part of The Connaught Automotive Research (CAR). They have looked closely at the information capacity numbers for a variety of camera parameters and the metric correspondence to machine vision performance.

Figure 1 shows the simulation chart used in the Deegan et al. research, which contains sets of four different sizes of simulated people (used as the test objects) and a slanted edge used to make the information capacity measurements. A series of images with various degradation levels of contrast, Gaussian blur, and Poisson noise were generated and evaluated using a YoloV10m model. The confidence interval of object detection was obtained as a performance measurement of the model.

Figure 1. A simulated test chart used for the Deegan et al. research that contains four specific sets of differently sized people
(Very large, Large, Medium, and Small) for quantifying the machine vision performance and a slanted square
with edges used for quantifying the traditional image quality metrics and new image information metrics.

Figure 2 shows the relationship between machine vision performance and a traditional image quality metric, MTF50. For this plot, the correlation of the model’s average confidence with MTF50 is low overall, which indicates the shortcoming of traditional metrics. Note that the data points with zero average confidence are the results for the small people category, which were too small to be detected by the machine vision model.

Figure 2. Machine Vision confidence versus MTF50.  for a test chart with four specific sizes of people
(Very large, Large, Medium, and Small). This graph shows that the machine vision model‘s confidence levels and MTF50
are poorly correlated. Note: The data points with zero average confidence are the results for the small people category,
which were too small to be detected by the machine vision model.

Figure 3 contains a similar comparison between machine vision model confidence and the new Shannon image information capacity metric (SIC in the figure). The correlation between the machine vision model confidence and the new information capacity is high. The green, yellow, red, and blue data points are for the very large, large, medium and small categories of people size, respectively. The small people (blue) were too small to be detected These sets of data indicate that information capacity can be used as a good metric for characterizing the machine vision that is part of AI.

Fig. 3. Machine Vision confidence versus Information Capacity.  for a test chart that has four specific sizes of people
(Very large, Large, Medium, and Small), this graph shows that the machine vision model‘s confidence levels
correlate well with the Shannon information capacity (SIC) for each grouping of people size detection
(with the exception of the detection of small people, which were too small to be detected by the machine vision model).

Imatest continues to work with our partners in academia and industry to further validate the image information metrics. In addition, we are leading the effort in the ISO/TC42 Photography standards body to bring these metrics into a new standard, ISO 23654 Image Information Metrics. This effort is currently in the new work initiative stage.

If you would like to collaborate with Imatest on the validation and standardization of image information metrics, reach out to us at infocap@imatest.com. You can also try out the image information metrics directly by using our Imatest analysis software, which has the metrics included in our newer releases. See more details at imatest.com/imaging/image-information-metrics/.

Future Work

• Complete the publication of the study that is summarized above.
• Study the correlation between detection confidence and additional image information metrics, including ideal observer SNR (SNRi) and Edge Location Sigma.
• Perform real-world camera tests instead of simulations. 
• Examine various filtering methods to determine if they improve object detection performance.
• Simulate or reproduce other failure modes in an imaging lab and consider their impact on object detection and image quality (IQ) metrics.

See Also

• Image Information Metrics (Imaging Page)
• Correlating the Performance of Computer Vision Algorithms with Objective Image Quality Metrics (H.Koren 2022)
• Image information metrics from Slanted edges: Instructions
• Image information metrics from Slanted edges: Equations and Algorithms
• Shannon information capacity from Siemens stars

The post Validating Information Metrics Correlation with Object Detection appeared first on Imatest.