Low sensitivity for orientation in texture similarity ratings

Abstract

Research on visual texture perception in the last decades was often devoted to segmentation and region segregation. In this report, I address a different aspect, that of texture identification and similarity ratings between texture fields with different texture properties superimposed. In a series of experiments, I noticed that certain feature dimensions were considered as more important for similarity evaluation than others. A particularly low ranking is given to orientation. This paper reports data from two test series: a comparison of color and line orientation and a comparison of two purely spatial properties, texture granularity (spatial frequency) and texture orientation. In both experiments, observers tended to ignore orientation when grouping texture patches for similarity and instead looked for similarities in the second dimension, color or spatial frequency, even across different orientations.

Introduction

Texture perception has been an important topic in vision research during the last 50 to 60 years. The early observation that some texture variations spontaneously segregate and others do not, even when the texture differences are recognized (Beck, 1966; Beck, 1982; Julesz, 1962; Julesz, 1975; Mayhew & Frisby, 1978; Olson & Attneave, 1970), has set off a large number of studies to explore, model, and explain this distinction. One particularly interesting feature dimension that provides segregation is orientation (Figure 1). The assumption that segmentation is based on “textons” (Julesz, 1981; Julesz, 1984) (i.e., a limited number of specific features) has led to a wide search for more such textons and to interesting discoveries (e.g., Enns, 1986; Victor, Rizvi, & Conte, 2019). But, several papers in the 1980s (Bergen & Adelson, 1988; Caelli, 1982; Fogel & Sagi, 1989; Turner, 1986; Voorhees & Poggio, 1988) showed that nearly all discovered textons (e.g., blobs, oriented lines, crossings, closure) can also be detected and distinguished by Gabor filters that select certain spatial frequencies and orientations. Texture segmentation could, thus, be explained by a combination of oriented spatial frequency filters, which resemble the properties of neurons in the early visual system. 

Figure 1.

 

Different issues of texture perception. (a, b) Texture segmentation provides the percept of segregating regions, but not all spatial variations generate that percept. Try to follow the patterned snakes through the texture patches. That should be easy in a (perceived segregation of orientation differences) but impossible in b (no spontaneous segregation of ۷s and ۸s). (c) Texture identification is important to determining to which of the partly hidden object surfaces a given texture patch may belong.

In addition to segmentation, the identification, classification, and evaluation of similarities are also important aspects of texture perception (e.g., Gurnsey & Fleet, 2001; Harvey & Gervais, 1981; Laws, Gale, & Leeson, 2003; Sun, St-Amand, Baker, & Kingdom, 2021)—for example, in computer-aided diagnosis (Li et al., 2022), material control and recognition (e.g., Balas & Greene, 2023), and in cognition, such as when observers must decide whether partly occluded texture fields belong to the surfaces of same or different objects (cf. Figure 1c). 

In a recent study, I noticed that features were differently weighed when observers had to rate texture patches for similarity. In pairwise comparisons, some features were considered important but others might have even been ignored. Here, I present two experiments from that study: the combination of color and orientation and that of two purely spatial properties, texture granularity (spatial frequency) and texture orientation. Observers were exposed to line textures that differed in line orientation and color (Figure 2) and to texture fields that differed in their oriented spatial frequency composition (Figure 4). In various test patterns, five texture patches were always generated so that three of them shared the same texture orientation but eventually differed in line color or texture grain, and three had the same color or grain but differed in orientation. These patches were randomly arranged in vertical or horizontal configurations, and observers were asked to indicate in which global configuration they found the patches looking more similar. 

Figure 2.

 

Example of a test pattern in Experiment 1. Which texture patches look more similar—the three vertical or the three horizontal ones? Most observers preferred similarities in color over similarities in orientation.

Although orientation is an efficient feature dimension in texture segmentation (Beck, 1982; Gheorghiu & Kingdom, 2012; Mayhew & Frisby, 1978; Norman, Haywood, & Kentridge, 2011; Nothdurft, 1991; Olson & Attneave, 1970; Victor, Thengone, & Conte, 2013), it seems to play only a minor role in similarity estimates when various texture features are combined. Color and texture granularity, on the other hand, generate strong impressions of surface identity, even across different texture orientations. 

Methods

This paper is an early report of two experiments from a larger study in which various features are compared (Nothdurft, 2025). In Experiment 1, line textures were used, in Experiment 2, noise patterns with defined spatial frequency content. 

Procedure

Test patterns such as those shown in Figures 2 and 4 were presented to eight observers with normal or corrected-to-normal visual acuity and non-impaired color vision (Ishihara test plates). All but one were students between the ages of 19 to 37 years (median, 25 years) and were paid for the time they spent in the experiment; one observer was the author (75 years). Observers had to indicate whether the three vertical or the three horizontal texture fields appeared more similar. They were encouraged to decide on their global perceptual impression of texture similarities in every new test pattern. 

During the tests, observers set relaxed in front of the monitor (73 cm away from the eyes) where stimuli occurred. Test patterns were shown for 2 seconds; during this time, observers could freely look around the pattern. Responses were entered on a computer keyboard. Shortly after the response, a new test pattern was shown. The tests were embedded in a larger set of experiments on various combinations of texture features; the data presented here were collected in two short test series of about 10 minutes each, which had been repeated several times in the course of the study. The tests included 18 and 32 test patterns, respectively, which were repeatedly presented in random sequence; all test conditions of an experiment were intermingled in one single run. 

Stimuli

Test patterns were computer generated (DOS VGA) and displayed on a monitor (Multiscan 17se II; Sony, Tokyo, Japan) at a refresh rate of 60 Hz. The individual texture fields were 4.6 deg × 4.6 deg; the entire test patterns covered an area of about 16 deg × 16 deg. 

Experiment 1 color patterns

Texture fields displayed red or green lines (size 0.3 deg × 0.1 deg) at one of two orientations (45° or 135°) which were arranged in a regular 7 × 7 raster (raster width, 0.65 deg), with a small jitter (0.1 deg) that was refreshed in every new pattern presentation. All lines in a single texture patch were identical; that is, they had the same orientation and color (see Figure 2). Red (27.2 cd/m²) and green colors were subjectively adjusted to iso-luminance by means of heterochromatic flicker minimization. Background luminance on the screen was 11.2 cd/m². Across conditions, color saturation was varied from 100% (maximally achieved isoluminant color settings on the monitor) down to 0% (white lines). The reduced color signals in intermediate saturation levels (50%, 20%, and 10%) were replaced by white so that all lines in the test displayed the same luminance. 

Experiment 2 noise patterns

Patterns were black and white in the luminance range of 5.5 to 67.8 cd/m² and displayed random noise with selected spatial frequency content (Figure 4). To generate such patterns, first the two-dimensional power distribution of each pattern was defined, and absent frequency components were set to zero (Figure 5). Noise patterns were then generated by randomizing the phases of all frequencies and computing the inverse two-dimensional Fourier transform. To avoid delays from pattern generation during the experiment, various patterns of each condition were computed beforehand and randomly loaded to the screen immediately before presentation. Four frequency bands of different granularity were used and tested in the experiment: (1) coarse texture granularity (frequency range, 0.86–1.73 cpd; mid-frequency, 1.30 cpd); (2) medium-coarse granularity (frequency range, 1.30–2.59 cpd; mid-frequency, 1.94 cpd); (3) medium-fine granularity (frequency range, 1.73–3.46 cpd; mid-frequency, 2.59 cpd); and (4) fine texture granularity (frequency range, 2.59–5.18 cpd; mid-frequency, 3.89 cpd). In each of these frequency bands, only an orientation range of ±10° around +45° or –45° from the vertical was activated in the patterns; all other orientations and frequency bands were set to zero. Figure 5 shows examples of the power distributions used. In the demo shown in Figure 4c, no orientation selection had been made for the central patch. Patterns were displayed at a luminance resolution of 64 levels over the full luminance range of the monitor. 

Results

Experiment 1. Color versus orientation

An examples of test patterns is shown in Figure 2. All (n = 8) but one observer spontaneously grouped texture patches for color, not for orientation, an impression that might be shared by the reader when looking at the example in Figure 2. To see whether the predominance of color would deteriorate with less saturated colors, patterns with fainter colors were also tested. But, even at reduced color saturations, most observers still preferred color over orientation when grouping the line patches for similarity (Figure 3). Only when the color was completely removed and white lines were shown instead (0% saturation, “noCOL” condition) did observers reliably group patterns for the similar line orientation, an indication that the orientation information was not ignored in the tests. Only one observer showed a partly opposite behavior by indicating that patterns with the same line orientation, independent of their colors, often looked similar to him. With highly saturated colors, however, he could not entirely ignore the strength of color grouping but gave indifferent responses with a similar proportion of color and orientation preferences (Figs. 2 and 3). 

Figure 3.

 

Mean performance of all eight observers in Experiment 1. Observers indicated which line patterns appeared more similar to them—the three vertical or the three horizontal ones (cf. Figure 2). Their performances were analyzed as preference for same line orientation (0%) or preference for same color (100%). All but one observer strongly preferred grouping for color, even when the color saturation was reduced and colors looked faint. When colored lines were replaced by white lines (0% color saturation, “noCOL”), all observers selected the patterns with identical line orientation as similar. Means and standard errors of the means are plotted.

Altogether, the data revealed a strong ranking of color and orientation when texture fields were evaluated for similarity. Line orientation was recognized but across the majority of observers was not considered as similarly important as line color when characterizing a texture field. The different ratings for color and no-color patches (four left-hand vs. the right-hand histogram bars in Figure 3) were each statistically significant (Wilcoxon signed-rank test; n = 8, T = 0, T_crit < 3; p <  0.05) and highly significant when pooled over all comparisons (n = 32, T = 0; T_crit < 94; p < 0.001). 

Experiment 2. Texture granularity versus orientation.

When differences in texture grain (spatial frequency) were strong enough to be seen, all observers rated the same-texture frequency patches as similar and were even willing to ignore the orientation differences between these patches. This is also the impression one may get from the examples in Figures 4a and 4b. The preference for spatial frequency over orientation (data points above 50% in Figure 6) was seen in all tested comparisons, provided the differences in texture grain were large enough to be recognized (distance from the open circle data point in each curve). It does not mean, however, that observers did not see and identify the texture orientations in different patches. When texture grains looked similar or were identical (open circles in Figure 6), observers instead used texture orientation for their similarity estimates. Thus, again, texture orientation was not ignored but only considered as less important than differences in texture granularity. 

Figure 4.

 

Illustration of test patterns in Experiment 2. Which texture patches in each example look more similar—the three vertical or the three horizontal ones? (a, b) When texture patches differed in granularity, observers selected the patches with similar granularity (horizontal in a, vertical in b), even across different texture orientations. (c) The percept of similarity, however, is not obtained with non-oriented textures of the same spatial frequency. The central patch in c displays all orientations at the coarser frequency band of the horizontal patches but is not perceptually grouped with them.

Figure 5.

 

Power functions of texture patterns as shown in Figure 4. (a) Schema of presentation. (b) Examples of power functions used in Experiment 2. Rows refer to patterns with the same frequency bands and columns to patterns with all (left) and two selected (middle and right) orientations. Absent frequency components are dark. Coarse and fine texture grains differed in the spatial frequency content (radial bands around the midpoint). Texture orientation is represented in small sections of these bands.

Figure 6.

 

Mean performance of observers in the similarity grouping tests of Experiment 2. When texture patches displayed the same frequency but different orientations, all observers grouped them for orientation (open circles, values near 0%). But, when patches also differed in granularity and could be grouped for either texture grain or texture orientation (see examples in Figure 4), all observers preferred granularity (spatial frequency, SF) over orientation (values toward 100%). Small grain differences were more difficult to see and produced intermediate responses. Different curves represent comparisons of various grain patterns with the finer texture at the open circle data point. Standard errors of the means are plotted if larger than symbol size.

Interestingly, the relative unimportance of texture orientation in similarity ratings does not indicate that orientation information is, in general, irrelevant in these matches. When the central texture patch in Figure 4a is replaced by a non-oriented texture patch of the same granularity (Figure 4c), no perceptual grouping occurs, and neither the vertical (different spatial frequency bands) nor the horizontal (same spatial frequency band) patches are seen as similar. 

Altogether, the data reveal a strong ranking of spatial frequency and orientation information when texture fields are evaluated for similarity. The orientation of the texture fields is recognized but not considered as important as the spatial frequency components for characterizing a texture patch. 

All filled data points on the curves in Figure 6 are significantly different from the according open-circle condition (Wilcoxon signed-rank test; n = 8; T = 0; T_crit < 3; p < 0.05) and highly significant when pooled along each curve with more than one measurement (n = 16 and 24, respectively; T = 0; T_crit < 8; p < 0.001) or when pooled for either the maximal or minimal grain differences across the curves (n = 24; T = 0; T_crit < 40; p < 0.001). That is, similarity ratings differed reliably between iso-frequency pairs (orientation preferred) and pairs with different frequencies (frequency preferred). 

Discussion

Experiments 1 and 2 presented two examples of multi-featured texture variations in which the feature dimension orientation did not turn out to be particularly important when texture patterns were rated for their apparent similarity. This may be unexpected, in particular when we recall that many earlier studies on texture segmentation were, in fact, based on similarity judgments and quite a few of them had found texture orientation as the distinguishing feature (e.g., Beck, 1966; Beck, 1982; Mayhew & Frisby, 1978; Olson & Attneave, 1970). But, the present findings are not in conflict with these studies; the discrepancy lies in the number of features to be looked at. When orientation was the only distinguishing parameter, as in the no-color condition of Experiment 1 (right-hand histogram bar in Figure 3) and the same-frequency matches in Experiment 2 (open circles in Figure 6), the present data confirm the importance of orientation in similarity matches. But, when other texture differences such as color or texture grain were also present and observers had to decide which features were the most important ones for similarity ratings, observers (and likely the reader, too) preferred color or texture granularity over texture orientation. 

Was orientation ignored?

One reviewer raised the suspicion that participants might have considered orientation in the tests as unimportant for the similarity judgments. If this had really been the case, the observed preferences for other parameters would not have been surprising. But this was not likely the case and observers had no reason to assume that orientation might have been less important than, for example, color or granularity. In fact, both experiments had a number of patterns in which color (Experiment 1) or spatial frequency (Experiment 2) differences were absent or small. With these test patterns, observers reliably grouped patches for the same orientation, indicating that this parameter was not ignored. Because any preset bias for one or the other feature dimension should have been a problem in the experiments, I had instructed all observers to rate texture patches for the apparent global texture similarities and strictly avoided any suggestions about individual feature dimensions. Intuitively, most observers looked equally at both features, except perhaps the one observer in Experiment 1 who apparently was more biased to orientation (but still made several similarity ratings in favor of color when the color patches were highly saturated). 

As the reader may verify in Figures 2 and 4, line and texture orientations are well seen and differences easily recognized. Nevertheless, orientation variations are likely more readily ignored in texture similarity ratings when compared with other feature dimensions. 

Previous studies

Variations in texture orientation or spatial frequency have been reported to segregate well (Bach, Schmitt, Quenzer, Meigen, & Fahle, 2000; Caelli, 1982; Caelli & Moraglia, 1985; Hunt & Meinhardt, 2021; Mayhew & Frisby, 1978), with no obvious threshold variations within each dimension (Caelli, Brettel, Rentschler, & Hilz, 1983). This has also been reported for patterns of differently oriented, colored lines (Callaghan, Lasaga, & Garner, 1986; Saarela & Landy, 2012). All tested feature variations alone segregate well, but when different feature dimensions are superimposed there may be interference. In congruent conditions (same texture borders in both feature dimensions), segregation is improved (Bach et al., 2000; Hunt & Meinhardt, 2021; Saarela & Landy, 2012); however, in conflicting conditions (different borders or irrelevant variations in one of the two dimensions), segmentation may be disturbed and one or the other feature can become dominant (Callaghan, 1989; Morgan, Adam, & Mollon, 1992; Pearson & Kingdom, 2001; Saarela & Landy, 2012). When comparing hue and orientation of lines in ambiguous displays, Callaghan (1989) found a quickly growing dominance of color with increasing hue variations, in agreement with the apparently overall dominance of color (with a large hue difference) in Experiment 1. Random color variations in a texture field may render the perceived segmentation of orientation differences difficult, except for dichromates, who cannot distinguish the superimposed colors (Morgan et al., 1992; Pearson & Kingdom, 2001). 

Feature ranking or variations in perceived strength?

This raises an important question. Could the observed ranking in similarity ratings not simply reflect differences in the perceived salience of texture features? Would it not seem obvious that perceptually strong differences not only segregate better (Nothdurft, 1992; Nothdurft, 1993) but also might perhaps dominate the similarity judgments between texture fields? In that case, rankings of color and spatial frequency over orientation might change or even reverse when the color differences or variations in texture grain become small and the difference in texture orientation is increased. To a small extent, this was indeed found. The strong preference for color was (slightly) reduced with low-saturated colors (Figure 3), and so was the preference for texture grain when the differences in texture granularity were small (Figure 6). But, these variations were small, although orientation differences already displayed the maximally possible contrast (90°). The conspicuity of orientation contrast cannot be further increased (Bergen & Julesz, 1983; Nothdurft, 1993). Faster and better segmentation of orientation differences is only obtained when line elements are oriented vertically or horizontally (Callaghan et al., 1986) or generally aligned (e.g., Gheorghiu & Kingdom, 2012; Harrison & Keeble, 2008; Nothdurft, 1992). Thus, even with maximal orientation contrast and small variations in color or texture grain, the latter were still emphasized in similarity matches. It is unlikely that these rating preferences could be explained alone with salience variations between texture fields. The examples in Figures 2 and 4 do not show obvious salience variations, and orientation differences were not more difficult to see than the superimposed color or spatial frequency variations. 

Similar rankings between color and orientation are also seen with segmentation. When different texture borders from color or orientation are closely superimposed, the color differences must be extremely weak to let the orientation-defined figures become well recognized (Nothdurft, 2025), despite the fact that orientation variations alone can easily be recognized (cf. Figure 1a). 

Performance asymmetries

In their study with Gabor signals, Caelli and Moraglia (1985) measured texture segmentation from differences in both orientation and frequency and found “that texture discrimination based upon frequency differences is partially inhibited when orientation variations of equal magnitudes are introduced in both textures” (p. 684). It would be interesting to compare this observation with the ranking of both parameters in the similarity ratings of the present study. In an illustration, the authors showed that a texture field made of Gabors at a spatial frequency different to that of Gabors in the surround, segregates less vividly when the Gabors vary in orientation. This is not necessarily in conflict with the present study, as the fact that the texture field still segregates demonstrates that texture similarities in spatial frequency can be grouped across orientations, exactly as shown in Figure 4. To verify that there exists a ranking of both parameters, we should look at texture fields that differ in orientation and see whether segmentation is also obtained when orientation similarities must be grouped across different spatial frequencies (which is not seen in Figure 4). The difference is illustrated in Figure 7. The central texture field is easily discriminated when it differs in both parameters (spatial frequency and orientation) from the surrounding background (Figure 7a). It also segregates when the texture regions differ in spatial frequency, and orientation is randomly varied so that similar frequencies have to be grouped over different orientations (Figure 7b). The segregation of the central field, however, is strongly reduced when regions differ in orientation, and texture similarities must be grouped across different spatial frequencies (Figure 7c). This is exactly what one should expect from the reported ranking of texture grain and texture orientation in Experiment 2. 

Figure 7.

 

Ranking of spatial frequency and orientation in texture perception. (a) When texture pairs differ in both spatial frequency and orientation, regions strongly segregate perceptually. (b) When texture pairs differ in spatial frequency and orientation is randomly varied, regions still segregate, although less strongly. Note that frequency similarities must be grouped across orientations for this percept. (c) When texture pairs differ in orientation and spatial frequency is varied, segregation is strongly reduced. The grouping of orientation differences across spatial frequencies is difficult if not impossible. (Parts a and b are modified from Caelli & Moraglia, 1985; the Gabor signals of the original paper are here replaced by similar looking black-and-white items.)

Why is orientation less important?

Given the perceived strength of orientation differences over many other spatial variations in texture segmentation (Figure 1), the finding that orientation is of so little importance in texture identification is surprising. Why does the visual system use orientation to detect texture borders and separate objects from the ground but is willing to ignore that information in the presence of other features when evaluating the similarity of surface textures? From an ecological point of view, that might be advantageous, however. In our real world, the perceived texture orientation of a surface is not constant over an object but strongly varies with its three-dimensional form (e.g., Bravo & Farid, 2001; Li & Zaidi, 2000) and with transformations of the object or the viewing process, distance, perspective, tilt or slant, and with object or head rotations (for many nice examples, see Aloimonos, 1988; Burge, McCann, & Geisler, 2016; Chen & Saunders, 2019; Chen & Saunders, 2020; Knill, 1998; Stevens, 1981). Variations in texture grain seem to be more robust under many such transformations (but not with slant and texture compression; cf. Chen & Saunders, 2020). 

The observation reported here also underlines another finding in texture research. Although the segmentation of oriented textures was originally explained by the similarity and dissimilarity of oriented items in different texture fields, it was later shown that texture borders are, in fact, detected from local orientation gradients (Landy & Bergen, 1991; Nothdurft, 1985; Nothdurft, 1991), so that borders could even occur between statistically identical texture fields (Nothdurft, 1985; Nothdurft, 1994). Thus, even in segmentation, the exact evaluation of the texture orientation in different surface regions would not be an important source of information about surface identities. 

Acknowledgments

Commercial relationships: none. 

Corresponding author: Hans-Christoph Nothdurft. 

Email: [e-mail]. 

Address: Visual Perception Laboratory (VPL) Göttingen, Göttingen 37085, Germany.