Natural motion signals, as a working definition, are those that are actually encountered by specific animals in the environment they normally operate in. The need to consider specific animals arises because the motion signals that are processed by a brain depend on the ethological and ecological context. Motion signals are determined by environmental motion, by the type and structure of locomotion of an animal, and by the visual topography of the world the animal operates in. We suggest that it is essential to consider natural motion signals in more detail, since they may reveal constraints that have shaped the evolution of motion detection and information processing mechanisms. The primary focus of this paper is to outline what needs to be considered and what is required to characterize the biologically relevant information content of the visual motion environment of an animal. In particular, we discuss the principal sources of image motion, critically assess the different ways of reconstructing, analysing and modelling natural motion signals, and briefly summarize current attempts to identify coding strategies, matched filters and optimization of neurones involved in processing visual information. We end with a survey of sensory and neural adaptations to show the multiple levels of processing at which motion filters have evolved under the influence of natural motion signals.

We present a review of perceptual image quality metrics and their application to still image compression. The review describes how image quality metrics can be used to guide an image compression scheme and outlines the advantages, disadvantages and limitations of a number of quality metrics. We examine a broad range of metrics ranging from simple mathematical measures to those which incorporate full perceptual models. We highlight some variation in the models for luminance adaptation and the contrast sensitivity function and discuss what appears to be a lack of a general consensus regarding the models which best describe contrast masking and error summation. We identify how the various perceptual components have been incorporated in quality metrics, and identify a number of psychophysical testing techniques that can be used to validate the metrics. We conclude by illustrating some of the issues discussed throughout the paper with a simple demonstration. (C) 1998 Elsevier Science B.V. All rights reserved.

We implemented and compared the performance of three lossy compression techniques (block DCT, wavelet, and Lapped orthogonal transform) for the compression of X-ray radiographs. The quantisation matrices for each technique was optimised based on a perceptual model developed by Watson [1,2]. We examined the ability of the perceptual model to identify the just noticeable difference compression point and compare the performance of the perceptual model to a commonly used metric, signal to noise ratio. We also examined the compression performance of the perceptually optimised techniques at the just noticeable difference level of compression. As a benchmark, the compression performance was compared to the SPIHT wavelet compression algorithm [3,4]. We found that the wavelet based perceptual model provided the most consistent estimate of the just noticeable difference threshold, the block DCT based technique was next, and signal to noise ratio was significantly worse. The compression performance was similar, at the just noticeable difference point, though block DCT provided slightly better compression than the wavelet technique, which in turn provided slightly better compression than the SPIHT wavelet compression algorithm.

In this paper, we compute quantization matrices which are tuned to individual images for JPEG compression of 12 bit radiographs. The quantization matrices were derived using a perceptual model, and tested in a set of psychophysical experiments to find the just-noticeable- difference (jnd) quality level between the original and lossy compressed images. Each of the images used in the study was compressed to 17 different quality levels, where 'quality level' is a perceptually meaningful metric. The results show that the technique can be used to provide selectable quality image compression of radiographs. We provide a list of recommended quantization matrices at various quality levels for images of this class (peripheral bone radiographs).

The precision of quantitative and subjective evaluations of phantom image quality has been studied. Twenty-seven images of the American College of Radiology (ACR) mammography accreditation phantom were acquired under different x-ray techniques and digitized. Several quantitative image quality measures were obtained from each image by analyzing microcalcification and nodule target objects in the phantom. All images were also scored subjectively by 8 observers, each of whom provided a count of the number of objects seen in each target class (fibrils, microcalcifications, and nodules). An analysis was performed to predict the subjective measurements from the quantitative measurements and to estimate their variabilities. It was found that the subjective measures could be well predicted by the quantitative measures and that the variance of the quantitative measures was significantly smaller than that of the subjective measure, by almost a factor of 10. The implication for the ACR accreditation program for mammography is that a substantial improvement is possible in the image quality evaluation process by performing computerized analysis of the phantom images in addition to subjective analysis.

We asked a number of readers to evaluate 28 images of the ACR phantom acquired under a broad range of conditions (varying kVp, mAs, grid, no-grid, scatter materials). The phantom contains three types of structures: fibrils, microcalcification groups, and masses. The evaluation was performed according to the standard ACR criteria (i.e. counting the number of visible structures). The resulting scores were averaged across readers to obtain the average number of fibers, masses, and microcalcification groups seen for each image. The images were digitized and analzyed to obtain values for the noise level, background pixel value, and the contrast of image structures. We then found the linear combination of the image measurements which best predicted the reader scores. The variablity of the reader scores and variablity of the computer measures were also analyzed. We found that the computer measures of image contrast provide a good prediction of observer scores, have much less variablity than the observer scores, are straightforward to obtain, and are reproducible.

We investigate the hypothesis that the early visual system efficiently codes natural time varying images, first by tracking part of the image, then by matching the spatiotemporal properties of the neural pathway to those of the tracked image. A representation for the time varying image is formulated which consists of two spatiotemporal components, a velocity field component and a stationary component. We show, using digitized sequences of natural images, that the spatiotemporal spectrum and other attributes of the image markedly differ before and after tracking. The temporal frequency bandwidth and velocity distribution of the velocity field component are diminished in the region of tracking and broaden with increasing eccentricity from this region. On the other hand, the spectrum of the stationary component is unaffected by tracking. Comparison of the properties of the tracked image to those of the M and P pathways suggests that each pathway transmits different attributes of the tracked image. A retinal architecture which varies with eccentricity also matches the properties of the tracked image.

The authors calculate the spatiotemporal power spectrum of 14 image sequences in order to determine the degree to which the spectra are separable in space and time and to assess the validity of the commonly used exponential correlation model. They expand the spectrum by a singular value decomposition into a sum of separable terms and define an index of spatiotemporal separability. as the fraction of the signal energy that can be represented by the first (largest) separable term. All spectra were found to be highly separable with an index of separability above 0.98. The power spectra of the sequences were well fit by a separable model, which corresponds to a product of exponential autocorrelation functions separable in space and time.