Spatial arrangement of line, edge and grating detectors revealed by subthreshold summation

Efficient Deep Learning Architecture for Detection and Recognition of Thyroid Nodules

Ultrasonography is widely used in the clinical diagnosis of thyroid nodules. Ultrasound images of thyroid nodules have different appearances, interior features, and blurred borders that are difficult for a physician to diagnose into malignant or benign types merely through visual recognition. The development of artificial intelligence, especially deep learning, has led to great advances in the field of medical image diagnosis. However, there are some challenges to achieve precision and efficiency in the recognition of thyroid nodules. In this work, we propose a deep learning architecture, you only look once v3 dense multireceptive fields convolutional neural network (YOLOv3-DMRF), based on YOLOv3. It comprises a DMRF-CNN and multiscale detection layers. In DMRF-CNN, we integrate dilated convolution with different dilation rates to continue passing the edge and the texture features to deeper layers. Two different scale detection layers are deployed to recognize the different sizes of the thyroid nodules. We used two datasets to train and evaluate the YOLOv3-DMRF during the experiments. One dataset includes 699 original ultrasound images of thyroid nodules collected from a local health physical center. We obtained 10,485 images after data augmentation. Another dataset is an open-access dataset that includes ultrasound images of 111 malignant and 41 benign thyroid nodules. Average precision (AP) and mean average precision (mAP) are used as the metrics for quantitative and qualitative evaluations. We compared the proposed YOLOv3-DMRF with some state-of-the-art deep learning networks. The experimental results show that YOLOv3-DMRF outperforms others on mAP and detection time on both the datasets. Specifically, the values of mAP and detection time were 90.05 and 95.23% and 3.7 and 2.2 s, respectively, on the two test datasets. Experimental results demonstrate that the proposed YOLOv3-DMRF is efficient for detection and recognition of thyroid nodules for ultrasound images.

Unaltered Perception of Suprathreshold Contrast in Early Glaucoma Despite Sensitivity Loss

Purpose

Glaucoma raises contrast detection thresholds, but our natural visual environment is dominated by high contrast that may remain suprathreshold in early to moderate glaucoma. This study investigates the effect of glaucoma on the apparent contrast of visible stimuli.

Methods

Twenty participants with glaucoma with partial visual field defects (mean age, 72 ± 7 years) and 20 age‐similar healthy controls (mean age, 70 ± 7 years) took part. Contrast detection thresholds for Gabor stimuli (SD, 0.75°) of four spatial frequencies (0.5, 1.0, 2.0, and 4.0 c/deg) were first measured at 10° eccentricity, both within and outside of visual field defects for participants with glaucoma. Subsequently, the contrast of a central Gabor was matched to that of a peripheral Gabor with contrast fixed at two times or four times the detection threshold. Data were analyzed by linear mixed modelling.

Results

Compared with controls, detection thresholds for participants with glaucoma were raised by 0.05 ± 0.025 (Michelson units, ± SE; P = 0.12) and by 0.141 ± 0.026 (P < 0.001) outside and within visual field defects, respectively. For reference stimuli at two times the detection contrast, matched contrast ratios (matched/reference contrast) were 0.16 ± 0.039 (P < 0.001) higher outside compared with within visual field defects in participants with glaucoma. Matched contrast ratios within visual field defects were similar to controls (mean 0.033 ± 0.066 lower; P = 0.87). For reference stimuli at four times the detection contrast, matched contrast ratios were similar across all three groups (P = 0.58). Spatial frequency had a minimal effect on matched contrast ratios.

Conclusions

Despite decreased contrast sensitivity, people with glaucoma perceive the contrast of visible suprathreshold stimuli similarly to healthy controls. These results suggest possible compensation for sensitivity loss in the visual system.

The processing of compound radial frequency patterns

Radial frequency (RF) patterns can be combined to construct complex shapes. Previous studies have suggested that such complex shapes may be encoded by multiple, narrowly-tuned RF shape channels. To test this hypothesis, thresholds were measured for detection and discrimination of various combinations of two RF components. Results show evidence of summation: sensitivity for the compounds was better than that for the components, with little effect of the components' relative phase. If both RF components are processed separately at the point of detection, they would combine by probability summation (PS), resulting in only a small increase in sensitivity for the compound compared to the components. Summation exceeding the prediction of PS suggests a form of additive summation (AS) by a common mechanism. Data were compared to predictions of winner-take-all, where only the strongest component contributes to detection, a single channel AS model, and multi-channel PS and AS models. The multi-channel PS and AS models were modelled under both Fixed and Matched Attention Window scenarios, the former assuming a single internal noise source for both components and compounds or different internal noise sources for components and compounds respectively. The winner-take-all and single channel models could be rejected. Of the remaining models, the best performing one was an AS model with a Fixed Attention Window, consistent with detection being mediated by channels that are efficiently combined and limited by a single source of noise for both components and compounds.

Estimates of edge detection filters in human vision

Edge detection is widely believed to be an important early stage in human visual processing. However, there have been relatively few attempts to map human edge detection filters. In this study, observers had to locate a randomly placed step edge in brown noise (the integral of white noise) with a 1/f² power spectrum. Their responses were modelled by assuming the probability the observer chose an edge location depended on the response of their own edge detection filter to that location. The observer's edge detection filter was then estimated by maximum likelihood methods. The filters obtained were odd-symmetric and similar to a derivative of Gaussian, with a peak-to-trough width of 0.1-0.15 degrees. These filters are compared with previous estimates of edge detectors in humans, and with neurophysiological receptive fields and theoretical edge detectors.

Right-Hemisphere Superiority in the Discrimination of Spatial Phase

Visual field differences have been investigated in various detection and discrimination tasks for simple sinusoidal gratings or for complex gratings composed of two sinusoids of spatial frequencies f and 3 f. Sinusoidal gratings were employed to evaluate contrast sensitivity, subthreshold summation effects, aftereffects of adaptation to a high-contrast grating, and spatial-frequency discrimination. The tasks with complex gratings were detection of the 3 f component in the presence of a high-contrast f component and spatial-phase discrimination. The stimuli were presented either in the left or in the right visual hemifield. The results indicate a lack of lateralization for detection and spatial-frequency discrimination of sinusoidal gratings, and for the bandwidth of subthreshold summation effects and adaptation aftereffects, whereas the detection of the 3 f component in the presence of a high-contrast f component, as well as spatial-phase discrimination of f + 3 f gratings, show a left-field advantage. This suggests a right-hemisphere superiority in the processing of spatial phase.

Independent Spatial-Frequency-Tuned Channels in Binocular Fusion and Rivalry

Monocular masking studies show that the visibility of a one-dimensional sinusoidal grating remains unchanged in the presence of masking noise filtered so as to contain spectral components that are at least two octaves away from the spatial frequency of the grating (Stromeyer and Julesz 1972). In the present study, random-dot stereograms were bandpass filtered in the two-dimensional Fourier domain, and masking noise of various spatial frequency bands was added to the filtered stereograms. Masking noise bands containing equally effective noise energy were selected such that their bands were either overlapping with the stereoscopic image spectrum or were two octaves distant. The first case resulted in binocular rivalry; however, in the second case stereoscopic fusion could be maintained in the presence of strong binocular rivalry owing to the masking noise. This finding indicates that spatial-frequency-tuned channels are not restricted to one-dimensional gratings but operate on two-dimensional patterns as well. Furthermore, these frequency channels are utilized in stereopsis and work independently from each other, since some of these channels can be in binocular rivalry while at the same time other channels yield fusion. The main binocular experiments are demonstrated.

Quantitative analysis of illusory movement: spatial filtering and line localization in the human visual system

A narrow bar or line (width around 1 arcmin) between two fields of which the luminances are sinusoidally and in counterphase modulated in time appears to make an oscillatory movement. It is possible to annihilate this illusory movement with a real movement and thus to analyze this phenomenon quantitatively. Confirming previous studies, the amount of illusory movement (amplitude typically 10 arcsec) was proportional to the modulation depth of the fields and inversely proportional to the line width and the line contrast. The amount of illusory movement increased with defocus, a lower mean luminance, and eccentricity. The experimental results could be explained by a model that includes a linear low-pass spatial filter. For a Gaussian spatial filter, the standard deviation as derived from the experimental results was 1.1 arcmin (1.0-1.3 arcmin) (median with range of four observers) for well-focused, photopic, foveal viewing. We explored various criteria for line localization in the model (extremes and zero-crossings of Gaussian derivatives).

A parametric study of fear generalization to faces and non-face objects: relationship to discrimination thresholds

Fear generalization is the production of fear responses to a stimulus that is similar—but not identical—to a threatening stimulus. Although prior studies have found that fear generalization magnitudes are qualitatively related to the degree of perceptual similarity to the threatening stimulus, the precise relationship between these two functions has not been measured systematically. Also, it remains unknown whether fear generalization mechanisms differ for social and non-social information. To examine these questions, we measured perceptual discrimination and fear generalization in the same subjects, using images of human faces and non-face control stimuli (“blobs”) that were perceptually matched to the faces. First, each subject’s ability to discriminate between pairs of faces or blobs was measured. Each subject then underwent a Pavlovian fear conditioning procedure, in which each of the paired conditioned stimuli (CS) were either followed (CS+) or not followed (CS−) by a shock. Skin conductance responses (SCRs) were also measured. Subjects were then presented with the CS+, CS− and five levels of a CS+-to-CS− morph continuum between the paired stimuli, which were identified based on individual discrimination thresholds. Finally, subjects rated the likelihood that each stimulus had been followed by a shock. Subjects showed both autonomic (SCR-based) and conscious (ratings-based) fear responses to morphs that they could not discriminate from the CS+ (generalization). For both faces and non-face objects, fear generalization was not found above discrimination thresholds. However, subjects exhibited greater fear generalization in the shock likelihood ratings compared to the SCRs, particularly for faces. These findings reveal that autonomic threat detection mechanisms in humans are highly sensitive to small perceptual differences between stimuli. Also, the conscious evaluation of threat shows broader generalization than autonomic responses, biased towards labeling a stimulus as threatening.

Orientation tuning in human colour vision at detection threshold

We measure the orientation tuning of red-green colour and luminance vision at low (0.375 c/deg) and mid (1.5 c/deg) spatial frequencies using the low-contrast psychophysical method of subthreshold summation. Orientation bandwidths of the underlying neural detectors are found using a model involving Minkowski summation of the rectified outputs of a bank of oriented filters. At 1.5 c/deg, we find orientation-tuned detectors with similar bandwidths for chromatic and achromatic contrast. At 0.375 c/deg, orientation tuning is preserved with no change in bandwidth for achromatic stimuli, however, for chromatic stimuli orientation tuning becomes extremely broad, compatible with detection by non-oriented colour detectors. A non-oriented colour detector, previously reported in single cells in primate V1 but not psychophysically in humans, can transmit crucial information about the color of larger areas or surfaces whereas orientation-tuned detectors are required to detect the colour or luminance edges that delineate an object's shape.

Edges, curvature, and primal sketches

Marr described two versions of the primal sketch: the basic image-processing level in human vision. In line with his broader view of how one should construct explanatory theories in vision, he provided some details of the computational theory for this stage, the algorithms used, and how they might be implemented in neural systems. In this paper I consider how Marr ideas have continued over the past 30 years. In this regard, I pay particular attention to three stages: locating edges; describing edge curvature; linking local edge segments into elongated contours.

Why the brain knows more than we do: non-conscious representations and their role in the construction of conscious experience

Scientific studies have shown that non-conscious stimuli and representations influence information processing during conscious experience. In the light of such evidence, questions about potential functional links between non-conscious brain representations and conscious experience arise. This article discusses neural model capable of explaining how statistical learning mechanisms in dedicated resonant circuits could generate specific temporal activity traces of non-conscious representations in the brain. How reentrant signaling, top-down matching, and statistical coincidence of such activity traces may lead to the progressive consolidation of temporal patterns that constitute the neural signatures of conscious experience in networks extending across large distances beyond functionally specialized brain regions is then explained.

Précis of Sensory Analysis

Sensory analysis is that initial, preconscious stage of perception at which primitive features (edges, temporal discontinuities, and periodicities) are picked out from the random fluctuations that characterize the physical stimulation of sensory receptors. Sensory analysis may be studied by means of signal-detection, psychometric-function, and threshold experiments, and Sensory Analysis presents a succinct, quasi-quantitative account of the phenomena revealed thereby. This account covers all five sensory modalities, emphasising the similarities between them.

A succinct account depends on identifying simple principles of wide generality, of which the most fundamental are that (a) sensory discriminations are differentially coupled to the physical stimuli and that (b) small stimuli are subject to a square-law transform which makes them less detectable than they would otherwise be. These two principles are established by comparisons between different configurations of two stimulus levels to be discriminated; they are realized within a simple physical-analogue model which affords certain low-level comparisons with neurophysiological observation. That physical-analogue model consists of a sequence of elementary operations on the stimulus constituting a stage of sensory processing. The concatenation of two or three stages in cascade accommodates an increased range of experimental phenomena, especially the detection of sinusoidal gratings.

This précis is organized in three parts: Part I surveys Sensory Analysis as economically as may be, beginning from the simplest, most fundamental ideas and working toward phenomena of increasing complexity. A rather shorter Part II reviews the most important alternative models addressed to some part or other of the phenomena surveyed. Finally, a very short Part III contributes some metatheoretic remarks on the function of a theory of sensory discrimination.

Can subthreshold summation be observed with the Ehrenstein illusion?

Subthreshold summation between physical target lines and illusory contours induced by edges such as those produced in the Kanizsa illusion has been reported in previous studies. Here, we investigated the ability of line-induced illusory contours, using Ehrenstein figures, to produce similar subthreshold summation. In the first experiment, three stimulus conditions were presented. The target line was superimposed on the illusory contour of a four-arm Ehrenstein figure, or the target was presented between two dots (which replaced the arms of the Ehrenstein figure), or the target was presented on an otherwise blank screen (control). Detection of the target line was significantly worse when presented on the illusory contour (on the Ehrenstein figure) than when presented between two dots. This result was consistent for both curved and straight target lines, as well as for a 100 ms presentation duration and unlimited presentation duration. Performance was worst in the control condition. The results for the three stimulus conditions were replicated in a second experiment in which an eight-arm Ehrenstein figure was used to produce a stronger and less ambiguous illusory contour. In the third experiment, the target was either superimposed on the illusory contour, or was located across the central gap (illusory surface) of the Ehrenstein figure, collinear with two arms of the figure. As in the first two experiments, the target was either presented on the Ehrenstein figure, or between dots, or on a blank screen. Detection was better in the dot condition than in the Ehrenstein condition, regardless of whether the target was presented on the illusory contour or collinear with the arms of the Ehrenstein figure. These three experiments demonstrate the ability of reduced spatial uncertainty to facilitate the detection of a target line, but do not provide any evidence for subthreshold summation between a physical target line and the illusory contours produced by an Ehrenstein figure. The incongruence of these results with previous findings on Kanizsa figures is discussed.

Narrow-band 1, 2, 3, 4, 8, 16, 24, 32, 48, 64, and 96 cycles/360 degrees angular frequency filters

We measured human frequency response functions for eleven angular frequency filters using a forced-choice procedure in a supra-threshold summation paradigm. Each of the eleven functions of 17 experimental conditions was measured 4-9 times among 12 observers. Results show that, for the arbitrarily selected filter phases, maximum summation effect occurred at test frequency for all filters. These results lead to the conclusion that there are narrow-band angular frequency filters operating in human visual system mostly through summation surrounded by inhibition at the specific test frequency ranges. Our previous suggestion (Simas and Santos, 2002), arguing that summation for the higher angular frequency filters should occur if background angular frequency contrast were set to a maximum of 5 times the test frequency threshold, was supported.

The structure of stereoscopic masking: Position, disparity, and size tuning

The masking effect of a Gaussian blob on detection of a Gaussian target was measured as a function of the position, disparity, width and polarity of the mask. The data reveal a large degree of disparity-specific masking that cannot be explained by the masking of its monocular constituents. At 5 degrees eccentricity, the masking range extends about +/-1 degrees around the lines of sight of the two eyes and 1-3 degrees in disparity, depending on the size of the test stimuli. The masking effects can be modeled as having three additive components, one that has a fixed disparity range and is polarity independent, one with a center/surround form keyed to both the disparity and the polarity of the mask, and one that derives from the monocular masking in each eye. Thus, the profound disparity interaction behavior is not limited to the simple monocular masking properties of the stimuli but reveals extensive connectivity across the disparity domain. Future models of disparity encoding will need to take these properties into account.

Effects of global shape on angle discrimination

Previous studies have been inconclusive as to whether angle discrimination performance can be predicted by the sensitivity of orientation discrimination mechanisms or by that of mechanisms specialised for angle coding. However, these studies have assumed that angle discrimination is independent of the shape of the object of which the angle is a part. This assumption was tested by measuring angle discrimination using angles that were parts of different triangular shapes. Angle discrimination thresholds were lowest when angles were presented in isosceles triangles (sides forming the angle were of identical length). Performance was significantly poorer when angles were presented in scalene triangles (sides of different lengths) and as much as three times worse when the sides forming the angle varied randomly in length between presentations. Comparing orientation discrimination for single lines with angle discrimination for different stimulus conditions (isosceles, scalene and random triangles) leads to conflicting conclusions as to the mechanisms underlying angle perception: line orientation sensitivity correctly predicts angle discrimination for random triangles, but underestimates angle acuity for isosceles triangles. The fact that performance in angle discrimination tasks is strongly dependant on the overall stimulus geometry implies that geometric angles are computed by mechanisms that are sensitive to global aspects of the stimulus.

Edges and bars: where do people see features in 1-D images?

There have been two main approaches to feature detection in human and computer vision--based either on the luminance distribution and its spatial derivatives, or on the spatial distribution of local contrast energy. Thus, bars and edges might arise from peaks of luminance and luminance gradient respectively, or bars and edges might be found at peaks of local energy, where local phases are aligned across spatial frequency. This basic issue of definition is important because it guides more detailed models and interpretations of early vision. Which approach better describes the perceived positions of features in images? We used the class of 1-D images defined by Morrone and Burr in which the amplitude spectrum is that of a (partially blurred) square-wave and all Fourier components have a common phase. Observers used a cursor to mark where bars and edges were seen for different test phases (Experiment 1) or judged the spatial alignment of contours that had different phases (e.g. 0 degrees and 45 degrees ; Experiment 2). The feature positions defined by both tasks shifted systematically to the left or right according to the sign of the phase offset, increasing with the degree of blur. These shifts were well predicted by the location of luminance peaks (bars) and gradient peaks (edges), but not by energy peaks which (by design) predicted no shift at all. These results encourage models based on a Gaussian-derivative framework, but do not support the idea that human vision uses points of phase alignment to find local, first-order features. Nevertheless, we argue that both approaches are presently incomplete and a better understanding of early vision may combine insights from both.

Flexibility of spatial averaging in visual perception

The classical receptive field (RF) concept-the idea that a visual neuron responds to fixed parts and properties of a stimulus-has been challenged by a series of recent physiological results. Here, we extend these findings to human vision, demonstrating that the extent of spatial averaging in contrast perception is also flexible, depending strongly on stimulus contrast and uniformity. At low contrast, spatial averaging is greatest (about 11 min of arc) within uniform regions such as edges, as expected if the relevant neurons have orientation-selective RFs. At high contrast, spatial averaging is minimal. These results can be understood if the visual system is balancing a trade-off between noise reduction, which favours large areas of averaging, and detail preservation, which favours minimal averaging. Two distinct populations of neurons with hard-wired RFs could account for our results, as could the more intriguing possibility of dynamic, contrast-dependent RFs.

Sistema visual humano: evidência psicofísica para filtros de freqüência angular baixa

O objetivo deste trabalho foi mensurar curvas de resposta ao contraste para os filtros de freqüências angulares de banda estreita de 1, 2, 3 e 4 ciclos/360º. Foram estimadas nove curvas para cada filtro com o método psicofísico de somação de resposta de supralimiar aliado ao método da escolha forçada. Tomaram parte neste experimento cinco participantes adultos com acuidade visual normal ou corrigida. Os resultados demonstraram somações máximas de limiar de contraste na freqüência de teste dos filtros de 1, 2, 3 e 4 ciclos/360º circundadas por inibições nas freqüências vizinhas às freqüências angulares de teste de cada filtro. Estes resultados são consistentes com a existência de filtros de freqüências angulares de banda estreita operando no sistema visual humano através do processo de somação ou inibição na faixa de freqüências angular baixa.

Radial frequency stimuli and sine-wave gratings seem to be processed by distinct contrast brain mechanisms

An assumption commonly made in the study of visual perception is that the lower the contrast threshold for a given stimulus, the more sensitive and selective will be the mechanism that processes it. On the basis of this consideration, we investigated contrast thresholds for two classes of stimuli: sine-wave gratings and radial frequency stimuli (i.e., j0 targets or stimuli modulated by spherical Bessel functions). Employing a suprathreshold summation method, we measured the selectivity of spatial and radial frequency filters using either sine-wave gratings or j0 target contrast profiles at either 1 or 4 cycles per degree of visual angle (cpd), as the test frequencies. Thus, in a forced-choice trial, observers chose between a background spatial (or radial) frequency alone and the given background stimulus plus the test frequency (1 or 4 cpd sine-wave grating or radial frequency). Contrary to our expectations, the results showed elevated thresholds (i.e., inhibition) for sine-wave gratings and decreased thresholds (i.e., summation) for radial frequencies when background and test frequencies were identical. This was true for both 1- and 4-cpd test frequencies. This finding suggests that sine-wave gratings and radial frequency stimuli are processed by different quasi-linear systems, one working at low luminance and contrast level (sine-wave gratings) and the other at high luminance and contrast levels (radial frequency stimuli). We think that this interpretation is consistent with distinct foveal only and foveal-parafoveal mechanisms involving striate and/or other higher visual areas (i.e., V2 and V4).

Separation of edge detection and brightness perception

When a low spatial frequency noise mask is superimposed onto a luminance staircase, the perceived brightness pattern is dramatically altered although the edges remain visible. We measured contrast thresholds for the edges and for the illusory scalloping (Chevreul-illusion), as a function of noise center spatial frequency. The masking tuning functions overlapped, but peaked at different spatial frequencies and contrast levels. The results suggest that perceived brightness is triggered only by the low spatial frequency components of the edges--the high spatial frequency components are not able to produce a brightness pattern.

Neural fine tuning during Vernier acuity training?

A superposition masking and summation to threshold paradigm was employed before and after unmasked Vernier acuity training to measure sensory changes of offset analysing mechanisms. Masking functions show a uniform downward translation after training and detection data reveal higher sensitivities to compound Gabor gratings in the post-test. These findings confirm the existence of learning related changes at early levels of information processing, but the results cannot be explained by neural fine tuning of offset analysing mechanisms. The data are consistent with the idea of task dependent broadening of orientation tuned mechanisms responsible for detecting small Vernier offsets.

Contrast discrimination with sinusoidal gratings of different spatial frequency

The detectability of contrast increments was measured as a function of the contrast of a masking or "pedestal" grating at a number of different spatial frequencies ranging from 2 to 16 cycles per degree of visual angle. The pedestal grating always had the same orientation, spatial frequency, and phase as the signal. The shape of the contrast-increment threshold versus pedestal contrast (TvC) functions depends on the performance level used to define the "threshold," but when both axes are normalized by the contrast corresponding to 75% correct detection at each frequency, the TvC functions at a given performance level are identical. Confidence intervals on the slope of the rising part of the TvC functions are so wide that it is not possible with our data to reject Weber's law.

The role of spatial frequency channels in letter identification

How we see is today explained by physical optics and retinal transduction, followed by feature detection, in the cortex, by a bank of parallel independent spatial-frequency-selective channels. It is assumed that the observer uses whichever channels are best for the task at hand. Our current results demand a revision of this framework: Observers are not free to choose which channels they use. We used critical-band masking to characterize the channels mediating identification of broadband signals: letters in a wide range of fonts (Sloan, Bookman, Künstler, Yung), alphabets (Roman and Chinese), and sizes (0.1-55 degrees ). We also tested sinewave and squarewave gratings. Masking always revealed a single channel, 1.6+/-0.7 octaves wide, with a center frequency that depends on letter size and alphabet. We define an alphabet's stroke frequency as the average number of lines crossed by a slice through a letter, divided by the letter width. For sharp-edged (i.e. broadband) signals, we find that stroke frequency completely determines channel frequency, independent of alphabet, font, and size. Moreover, even though observers have multiple channels, they always use the same channel for the same signals, even after hundreds of trials, regardless of whether the noise is low-pass, high-pass, or all-pass. This shows that observers identify letters through a single channel that is selected bottom-up, by the signal, not top-down by the observer. We thought shape would be processed similarly at all sizes. Bandlimited signals conform more to this expectation than do broadband signals. Here, we characterize processing by channel frequency. For sinewave gratings, as expected, channel frequency equals sinewave frequency f(channel)=f. For bandpass-filtered letters, channel frequency is proportional to center frequency f(channel) proportional, variantf(center) (log-log slope 1) when size is varied and the band (c/letter) is fixed, but channel frequency is less than proportional to center frequency f(channel) proportional, variantf(center)(2/3) (log-log slope 2/3) when the band is varied and size is fixed. Finally, our main result, for sharp-edged (i.e. broadband) letters and squarewaves, channel frequency depends solely on stroke frequency, f(channel)/10c/deg=(2/3), with a log-log slope of 2/3. Thus, large letters (and coarse squarewaves) are identified by their edges; small letters (and fine squarewaves) are identified by their gross strokes.

Lateral interactions: size does matter

Usually a high-contrast, co-local mask increases contrast threshold (inhibition). Interestingly, a laterally displaced mask (flanker) can facilitate contrast detection (Vision Research 33 (1993) 993; 34 (1994) 73). When spatial scaling of these flanker effects was implied, stimulus bandwidth was confounded with spatial frequency (lambda(-1)). Under conditions where at lower spatial frequencies, the size (standard deviation, sigma) of the Gabor patch was smaller (sigma<lambda) than higher spatial frequencies (sigma=lambda), the effect appeared scale invariant. We replicated the original results for all conditions. However, when Gabor size was fixed (sigma=lambda), facilitation changed with spatial frequency (range 2--13 cycles/deg). When Gabor size was varied (sigma=0.5-2 lambda), usually the combination of larger patch sizes and lower spatial frequencies caused inhibition. We were unable to find any conditions that demonstrated spatial scaling. The size, both lambda and sigma, of both stimulus and flankers, influenced contrast threshold. Also, facilitation reduced as contrast of the flankers was reduced to detection threshold. Some facilitation was apparent with sub-threshold flankers. These results need to be reconciled with current models of lateral interactions.

Narrow-band 1, 2, 3, 4, 8, 16 and 24 cycles/360 degrees angular frequency filters

We measured human frequency response functions for seven angular frequency filters whose test frequencies were centered at 1, 2, 3, 4, 8, 16 or 24 cycles/360 degree using a supra-threshold summation method. The seven functions of 17 experimental conditions each were measured nine times for five observers. For the arbitrarily selected filter phases, the maximum summation effect occurred at test frequency for filters at 1, 2, 3, 4 and 8 cycles/360 degree. For both 16 and 24 cycles/360 degree test frequencies, maximum summation occurred at the lower harmonics. These results allow us to conclude that there are narrow-band angular frequency filters operating somehow in the human visual system either through summation or inhibition of specific frequency ranges. Furthermore, as a general result, it appears that addition of higher angular frequencies to lower ones disturbs low angular frequency perception (i.e., 1, 2, 3 and 4 cycles/360 degree), whereas addition of lower harmonics to higher ones seems to improve detection of high angular frequency harmonics (i.e., 8, 16 and 24 cycles/360 degree). Finally, we discuss the possible involvement of coupled radial and angular frequency filters in face perception using an example where narrow-band low angular frequency filters could have a major role.

Visual summation of luminance lines and illusory contours induced by pictorial, motion, and disparity cues

Illusory contours where no contrast exists in the image can be seen between pairs of spatially separate but aligned inducing real contours defined either by pictorial cues (luminance contrasts or offset gratings), kinetic contrast, or binocular disparity contrast. In previous studies it has been shown that the detection of a thin luminous line is facilitated when the line is superimposed on illusory contours and the inducing flanking elements are defined by luminance contrast. By using a spatial forced-choice technique I show that luminous lines summate with illusory contours induced by luminance contrast, offset gratings, motion contrast, and disparity contrast when the line is superimposed on the illusory contour. Control experiments show that the positional cues, offered by the inducing contours, are unable to account for these results. It is suggested that real luminous lines or edges and illusory contours activate common neural mechanisms in the brain irrespectively of the stimulus attributes that induce the illusory contour.

Multiple processes mediate flicker sensitivity

By systematically manipulating the luminance of a flickering spot and the area immediately surrounding it, we investigated why thresholds from flickering stimuli that cause a change in average luminance are elevated relative to those from stimuli with no luminance change. Threshold elevation resulted from local light adaptation and from temporal-frequency-specific interactions between the spot and its surround: at low frequencies, the contrast between the spot and the surround elevated thresholds, whereas at high frequencies, dark adaptation within the surround elevated thresholds. Our findings suggest that two common ways of determining temporal sensitivity may give markedly different outcomes.

Human observer detection experiments with mammograms and power-law noise

We determined contrast thresholds for lesion detection as a function of lesion size in both mammograms and filtered noise backgrounds with the same average power spectrum, P(f)=B/f3. Experiments were done using hybrid images with digital images of tumors added to digitized normal backgrounds, displayed on a monochrome monitor. Four tumors were extracted from digitized specimen radiographs. The lesion sizes were varied by digital rescaling to cover the range from 0.5 to 16 mm. Amplitudes were varied to determine the value required for 92% correct detection in two-alternative forced-choice (2AFC) and 90% for search experiments. Three observers participated, two physicists and a radiologist. The 2AFC mammographic results demonstrated a novel contrast-detail (CD) diagram with threshold amplitudes that increased steadily (with slope of 0.3) with increasing size for lesions larger than 1 mm. The slopes for prewhitening model observers were about 0.4. Human efficiency relative to these models was as high as 90%. The CD diagram slopes for the 2AFC experiments with filtered noise were 0.44 for humans and 0.5 for models. Human efficiency relative to the ideal observer was about 40%. The difference in efficiencies for the two types of backgrounds indicates that breast structure cannot be considered to be pure random noise for 2AFC experiments. Instead, 2AFC human detection with mammographic backgrounds is limited by a combination of noise and deterministic masking effects. The search experiments also gave thresholds that increased with lesion size. However, there was no difference in human results for mammographic and filtered noise backgrounds, suggesting that breast structure can be considered to be pure random noise for this task. Our conclusion is that, in spite of the fact that mammographic backgrounds have nonstationary statistics, models based on statistical decision theory can still be applied successfully to estimate human performance.

Perceptual learning for a pattern discrimination task

Our goal was to differentiate low and mid level perceptual learning. We used a complex grating discrimination task that required observers to combine information across wide ranges of spatial frequency and orientation. Stimuli were 'wicker'-like textures containing two orthogonal signal components of 3 and 9 c/deg. Observers discriminated a 15% spatial frequency shift in these components. Stimuli also contained four noise components, separated from the signal components by at least 45 degrees of orientation or approximately 2 octaves in spatial frequency. In Experiment 1 naive observers were trained for eight sessions with a four-alternative same-different forced choice judgment with feedback. Observers showed significant learning, thresholds dropped to approximately 1/3 of their original value. In Experiment 2 we found that observers showed far less learning when the noise components were not present. Experiment 3 found, unlike many other studies, almost complete transfer of learning across orientation. The results of Experiments 2 and 3 suggest that, unlike many other perceptual learning studies, most learning in Experiment 1 occurs at mid to high levels of processing rather than within low level analyzers tuned for spatial frequency and orientation. Experiment 4 found that performance was more severely impaired by spatial frequency shifts in noise components of the same spatial frequency or orientation as the signal components (though there was significant variability between observers). This suggests that after training observers based their responses on mechanisms tuned for selective regions of Fourier space. Experiment 5 examined transfer of learning from a same-sign task (the two signal components both increased/decreased in spatial frequency) to an opposite-sign task (signal components shifted in opposite directions in frequency space). Transfer of learning from same-sign to opposite-sign tasks and vice versa was complete suggesting that observers combined information from the two signal components independently.

Apparent fineness of stationary compound gratings

Patterns consisting of the sum of a sinusoidal grating and its second spatial harmonic have an apparent spatial fineness, or periodicity, that is about halfway between the two component spatial frequencies. There are also phase dependent modulations of the apparent fineness about the mean fineness shift. Covariance between individuals' phase dependent fineness shifts indicates the presence of four spatial phase channels. The apparent fineness effects, and the putative phase channels, may both be a product of a local, linear, analysis of spatial frequency content. Illusory second harmonics, as generated in the spatial frequency doubling illusion, also change apparent fineness.

The interaction of first- and second-order cues to orientation

The visual system is sensitive to orientation information defined both by first-order (luminance) and by second-order (texture) cues. We consider how these orientation cues are computed and how they affect one another. We measured the perceived orientation of the first and second-order components of Gabor patches (the carrier and envelope, respectively) and report a dependence of the perceived orientation of each on the orientation of the other, and on the spatial frequency of the carrier. Fixing the carrier orientation near that of the envelope interferes with envelope orientation judgements. This interference is reduced by adding a small (subthreshold) rotation to the carrier indicating that the site of interference is early. When the gross relative orientation of carrier and envelope is varied, the carrier appears systematically tilted towards the envelope. However, provided envelope and carrier are separated by more than approximately 10 degrees, the perceived envelope orientation appears tilted away from the carrier. The size of these effects increases with decreasing carrier spatial frequency, and with increasing exposure duration. When the envelope and carrier are both non parallel and non-perpendicular Fourier energy is distributed asymmetrically across orientation. We demonstrate that, for a channel-based orientation code, this asymmetry induces a shift in mean orientation that is sufficient to explain illusory tilting of carriers. The illusory tilting of the envelope, as a function of carrier orientation and spatial frequency, demonstrates that human ability to demodulate contrast information is far from ideal and cannot be explained by existing two-stage filter-rectify-filter models. We propose that illusory tilting of the envelope is due to selective connectivity between first- and second-stage filters whose purpose is to dissociate the type of image structure producing each class of cue.

Evolving concepts of spatial channels in vision: from independence to nonlinear interactions

By the 1960s it was evident from neuroanatomy that there were extensive recurrent interactions, both excitatory and inhibitory, among visual cortical neurons. Nevertheless, the psychophysical discovery of 'spatial-frequency channels' gave rise to a decade in which parallel, independent channels were thought to subserve early spatial vision. Recent work, however, has clearly demonstrated that early visual channels do not perform a Fourier or wavelet decomposition of the image. Instead, they interact through a variety of nonlinear pooling mechanisms. Such nonlinear interactions perform important computations in texture perception, stereopsis, and motion and form vision.

Extraction of perceptually salient contours by striate cortical networks

We present a cortical-based model for computing the perceptual salience of contours embedded in noisy images. It has been suggested that horizontal intra-cortical connections in primary visual cortex may modulate contrast detection thresholds and pre-attentive "pop-out". In our model, horizontal connections mediate context-dependent facilitatory and inhibitory interactions among oriented cells. Strongly facilitated cells undergo temporal synchronization; and perceptual salience is determined by the level of synchronized activity. The model accounts for a range of reported psychophysical and physiological effects of contour salience. In particular, the model proposes that intrinsic properties of synchronization account for the increased salience of smooth, closed contours. Application of the model to real images is demonstrated.

Extraction of signatures from check background based on a filiformity criterion

Extracting a signature from a check with a patterned background is a thorny problem in image segmentation. Methods based on threshold techniques often necessitate meticulous postprocessing in order to correctly capture the handwritten information. In this study, we tackle the problem of extracting handwritten information by means of an intuitive approach that is close to human visual perception, defining a topological criterion specific to handwritten lines which we call filiformity. This approach was inspired by the existence in the human eye of cells whose specialized task is the extraction of lines. First, we define two topological measures of filiformity for binary objects. Next, we extend these measures to include gray-level images. One of these measures, which is particularly interesting, differentiates the contour lines of objects from the handwritten lines we are trying to isolate. The local value provided by this measure is then processed by global thresholding, taking into account information about the whole image. This processing step ends with a simple fast algorithm. Evaluation of the extraction algorithm carried out on 540 checks with 16 different background patterns demonstrates the robustness of the algorithm, particularly when the background depicts a scene.

Contrast sensitivity to angular frequency stimuli is higher than that for sinewave gratings in the respective middle range

This study compares contrast thresholds for sinewave gratings, or spatial frequencies (1/CSF) with contrast thresholds for angular frequencies (1/aCSF) and for radial frequencies, or Jzero targets (1/rCSF). Observers had to differentiate between one of these frequency stimuli and a stimulus at mean luminance within a forced-choice procedure. All measurements were made with the same equipment, methods and subjects. Our results show higher sensitivity to, or lower thresholds for, angular frequencies when compared to either sinewave gratings or Jzero targets. Contrast values in arbitrary units, in the lower threshold range for angular frequencies, were about half those required to differentiate sinewave gratings from mean luminance in its most sensitive range.

Matched filtering in the visual system of the fly : large monopolar cells of the lamina are optimized to detect moving edges and blobs

At high levels of ambient light, large monopolar cells (LMCS) display spatially antagonistic receptive fields and a biphasic response to a brief flash of light from an axially positioned point source. In low ambient light the response becomes monophasic everywhere within the receptive field. Using the theory of matched filters, we infer that the LMCS are optimal for the detection of moving edges at high light levels, and for ‘blobs’ in low ambient light. The spatio-temporal properties predicted by the theory are in agreement with experimental observation. At high light levels, the strong temporal inhibition, the weak, diffuse lateral inhibition, and the non-separability of the receptive field in space and time are all properties that promote the sensitivity to a moving edge. At low light levels, the lack of spatial or temporal antagonism enhances the sensitivity to a blob. Our hypothesis is reinforced by the observation that flies tend to walk toward the edges of a broad, dark vertical stripe at high light levels, but uniformly toward all regions within the stripe in low ambient light.

A test for multiplication in insect directional motion detectors

The H1 neuron is a directionally sensitive motion-detector neuron with a large field that is fed by many high-resolution motion detectors in the fly optic lobe. As a stimulus pattern for it we used a random pattern of 50% bright and 50% dark squares on an oscilloscope screen. When this pattern is jumped by a small increment the HI neuron gives a directional response. When the jump is greater than one pixel on the screen the response falls and becomes non-directional because jump direction can no longer be inferred. When the contrast is reversed at the jump, the response is the same for both directions, and is the same as when the contrast is reversed without motion. For the motion receptors this represents a nondirectional ‘on’ or ‘off’ response. The result is discussed with reference to theories of motion perception.

Psychophysical end-stopping associated with line targets

Increment threshold for a small (e.g. 1' x 5') line target superimposed on backgrounds of various shapes and sizes was measured to provide a detailed map of the spatial interactions about line targets. This modified "Westheimer paradigm" indicated sensitization in the length direction as well as in the width direction around the line target. The effect of the adaptation field summed over an elongated, end-tapered central region, and showed strong end-zone antagonism beyond the ends of the elongated summation area, as well as flank antagonism to the sides. Secondary disinhibitory and inhibitory areas outside of the antagonistic surround were also demonstrated. When length of the test line was varied, the length of the summation region increased concomitantly, while the length of the end-zones remained fixed. End-zone antagonism was slightly weaker at oblique orientations. These results demonstrate a perceptual analog to neurophysiological end-stopping, and suggest a multilobed y-dimension weighting profile appropriate for models of spatial visual abilities.

Neurophysiological evidence for contrast dependent long-range facilitation and suppression in the human visual cortex

Long-range spatial interactions in human visual cortex were explored using a lateral masking paradigm. Visual evoked potentials (VEPs) elicited by a Gabor signal presented in isolation or in the presence of two flanking high-contrast Gabor signals (masks) were measured. Response amplitude and phase were recorded for a vertically oriented test, for horizontal and vertical masks and for combinations of vertical tests and vertical or horizontal masks. The amplitudes and phases of the test alone and mask alone responses were added coherently to predict the amplitude for collinear and orthogonal lateral masking conditions. Additivity failures were taken as evidence for neural interactions. At a target-to-mask distance of 2 deg, VEP amplitude exceeded the linear prediction for test contrasts in the range of 8-16% for the collinear, co-axial target/mask combination. Measured response phase also led predicted response phase over the same range of contrast. The VEP amplitudes were less than the linear prediction in the orthogonal target/mask combination and measured response phase lagged the predicted phase. Significant facilitation occurred with collinear test/mask combinations up to at least 3 deg of separation (nine wavelengths). Co-oriented, but non-collinear test/mask combinations (oblique test and mask, horizontal test and mask) did not produce facilitation. Contrast gain thus appears to be set over considerable distances in a configuration-specific fashion.