Novel software performs image-quality evaluation
--ANDREW WILSON
Imaging-system developers have a vested interest in assessing the quality of captured images. Such assessments are needed for testing or optimizing imaging systems under development, checking the quality control of operating systems, and judging the comparative quality of images. In fingerprint-identification systems, for example, the Federal Bureau of Investigation uses a sine-wave target to measure the spatial frequency response (or modulation transfer function [MTF]) of image scanners to provide an overall measure of scanner quality (see Vision Systems Design, June 1997, p. 12).
"Quantitative measures of image quality are ideally superior to subjective measures because they are less variable and labor-intensive and produce results more quickly," says Norman B. Nill, lead imagery engineer at MITRE Corp. (Bedford, MA; [email protected]). "But in attempting to apply such quantitative measurements, it often happens that only the acquired images are available--there are no test images, reference images, or before/after images, or any capability to image test targets," he says.
For these situations, MITRE has developed digital image-quality-measurement (IQM) software that does not require imaging a specific pattern or target, a constant scene, or a reference image, nor does it rely on prediction of quality. Available on the Web at www.mitre.org/technology/mtf, these IQMs are obtained directly from computations on digital images presented to the software, without the need to search for or select specific scene components, such as edges. Thus, system de velopers can determine if reconnaissance imagery presents sufficient quality to fulfill given intelligence tasks or if a diagnosis can be performed from acquired x-ray images.
MITRE`s IQM software, developed with support from the Electronic Systems Center of the US Air Force, relies on an invariance of shape of a scene`s spatial frequency power spectrum, from scene to scene. "This spectrum shape invariance allows quality assessment directly from acquired images and eliminates the need for test targets in many situations," says Nill.
In operation, the IQM software first computes the power spectrum using a fast-Fourier-transform technique. Normalizations for digital image brightness and size are then incorporated into the power spectrum as is a perspective-controlled, directional scale factor that brings scale dependence back into the scale-independent power spectrum. The power spectrum is then filtered for noise and the human visual system. Studies have demonstrated that the visual system can be modeled with a spatial frequency response function. This function is then applied as a filter to the image spectrum, producing IQM values that closely track the human perceptual response to quality changes between images. Additionally, various image degradations can be detected and identified from information contained within the power spectrum itself, and the software incorporates a number of these degradation detectors, which are reported along with the quality value.
"Image quality is still an elusive quantity, because it can mean many different things," says Nill. "One would not expect reconnaissance image quality for military intelligence interpretability purposes to have the same meaning as fashion photog- raphy image quality," he says. "And although it is doubtful that any one measure of quality will ever encompass all of the definitions of image quality, most of these definitions include image sharpness, detail rendition, and contrast as major components of quality," Nill adds. These are the components that are measured by MITRE`s power-spectrum-based IQM.
