Region- and edge-based configurational effects in texture segmentation

Transcription

1 Vision Research 47 (2007) Region- and edge-based configurational effects in texture segmentation Enrico Giora *, Clara Casco Department of General Psychology, University of Padua, Via Venezia 8, Padua, Italy Received 12 April 2006; received in revised form 13 January 2007 Abstract We have found a new configurational effect in texture segmentation. In addition to collinear facilitation at the edge, this effect results from contextual modulation within the texture-region, i.e. from texels not abutting the segmented edge. The largest facilitation was found when two conditions were fulfilled: (i) elements along the edge were parallel to the edge and collinear, (ii) elements in the texture-region were also collinear but non-parallel to the edge. We show that this facilitation occurs when there are groups of different orientation from the edge in the texture-region. We suggest two possible underlying mechanisms: either a region-based process that links collinear iso-oriented elements and locates the edge when the orientation changes, or else second-order filters tuned to orientation differences rather than orientation per se. Ó 2007 Elsevier Ltd. All rights reserved. Keywords: Texture segmentation; Lateral interactions; Grouping by collinearity; Energy models 1. Introduction * Corresponding author. Fax: address: enrico.giora@unipd.it (E. Giora). One of the first processing steps on the path to perception is the segregation of objects from the background. This operation has often been studied using texture images where, in the absence of mean luminance differences, a given texture-region can be segmented on the basis of discontinuities of some basic dimension such as orientation, spatial-frequency or motion. The most popular accounts of texture segmentation are edge-based segregation models (see Landy & Graham, 2004; for a review). Briefly, these models predict that texture segmentation results from a non-linear transformation of the output of local spatial filters, followed by a 2ndorder spatial filtering to enhance activity at the texturedefined contours, where the local filter response changes. According to these models, edge-based segregation is thought to result from both enhanced signal processing at the texture-edge and local inhibitory activity in the texture-region, where filter response is weak (Malik & Perona, 1990; Sagi, 1991, p. 406). However, the hypothesis of inhibition is difficult to conciliate with the phenomenological evidence that local properties of texture-region can be still salient after segmentation. Indeed, when observing a pair of zebras, say, we perceive a pair of zebra coats, not just the boundary between their bodies, and this can only result from textureregion information, with no feature gradient (Ben Shahar, 2006). Roelfsema, Lamme, Spekreijse, & Bosch (2002) attempt to accommodate phenomenology with visual processing. They propose that during texture segregation, locations where the properties of texture-elements change abruptly are assigned to boundaries, whereas image regions that are relatively homogeneous are not inhibited, but perceived as texture-regions by grouping elements together. Moreover, psychophysical data strongly suggest that the properties of texture-regions are perceived by mechanisms different from those responsible for extracting texture edges (Ariely, 2001; Parkes, Lund, Angelucci, Solomon, & Morgan, 2001). Lee (1995) and Roelfsema et al. (2002) suggested that edge- and region-based mechanisms operate at different levels of processing /$ - see front matter Ó 2007 Elsevier Ltd. All rights reserved. doi: /j.visres

2 880 E. Giora, C. Casco / Vision Research 47 (2007) However, there are data suggesting that not only are textures perceived, but they produce contextual effects in segmentation. A well known contextual effect is a facilitation when texels are collinear and parallel to the edge (Caputo & Casco, 1999; Casco, Campana, Grieco, & Fuggetta, 2004; Casco, Grieco, Campana, Corvino, & Caputo, 2005; Nothdurft, 1992; Olson & Attneave, 1970; Wolfson & Landy, 1995). The underlying physiological mechanism (see Lamme, 2004; for a review) is based on the familiar phenomenon of enhancement of neuronal firing rate resulting from contextual influences from outside the receptive field. These contextual influences affect several visual tasks in addition to texture segmentation. For example, they reduce contrast threshold for a single target bar (Polat & Sagi, 1993, 1994) and enhance detection of contours embedded in background noise (Field, Hayes, & Hess, 1993; Hess & Dakin, 1997; Li & Gilbert, 2002). An account for contextual effects from the region in edge segmentation is proposed by region-based models. According to these models, the visual system treats neighbouring texture-regions as belonging to the same texture if they are similar enough. In this way, individual elements are grouped by spreading neural activity emanating from highly stimulated detectors, and the edge is detected when linking operations interrupt because the output of local filters changes (Caelli, 1985). Interestingly, since linking between collinear elements is preferred, edge segmentation by such a mechanism results from spreading of activity in the direction non-parallel to the edge and cannot be assimilated to the mechanism accounting for the facilitation produced by collinear elements parallel to the texture-edge. An alternative approach conciliates the contextual effects with edge-based models of texture segmentation. Wolfson & Landy (1999), for example, suggest that reduced detection of a target element oriented differently from the surrounding region, when the background surrounding the region is iso-oriented with the target (Caputo, 1996), could depend on inhibitory connections between orientation and spatial-frequency selective linear-filters. This explanation is, however, local and does not account for facilitatory contextual effects resulting from grouping in the texture-region. Thus, contextual influences from elements parallel to the edge are well compatible with edge-based models (as shown by Wolfson & Landy, 1995) and can modulate either facilitating or inhibiting edge-based segregation. On the other hand, contextual influences from the texture-region are instead taken in account by a region-based mechanism, predicting that a texture-edge is detected not explicitly but rather implicitly when the growing of two different texture-regions causes their interaction. In this case, facilitation should occur when collinear elements are nonparallel to the texture-edge. Our experiments were designed to investigate the interaction between region- and edge-based mechanisms. We predicted that, if texture-edge segmentation depended on region-based analysis, then contextual influences from the texture-region should affect the saliency of the edge. To test this hypothesis we checked how the texture overall-orientation affected the discrimination of the edge. In order to distinguish between the effect of collinearity at the edge and region-based effects, we arranged the edge to segment a larger texture-region from a narrow one. Consequently, when elements in the narrow region were parallel to the edge they were also collinear, whereas collinear elements in the larger texture were either parallel or orthogonal to the edge. Results show a new facilitatory configurational effect resulting from grouping by collinearity in the textureregion, and independent of collinear facilitation at the edge. 2. Methods 2.1. Stimuli Stimuli were generated by using a VSG 2/3 Cambridge Research System graphic card with 12-bit luminance resolution and displayed on a gamma-corrected Sony Triniton monitor with a resolution of pixels refreshed at 100 Hz. Observers viewed the stimuli in a dark room at 57 cm viewing distance. In all experiments, we used textures composed of 8 8 ( deg of visual angle) matrices of cosine-phase [even] Gabor-elements with circular support or envelope. Each Gabor-patch was defined as a sinusoidal-modulated carrier with a wavelength [k] of.31 deg (spatial-frequency of 3.2 cycles/deg) multiplied by a Gaussian envelope with standard deviation [r] of.19 deg. Centre-to-centre elements distance was equal to 3.66 k. Mean luminance of a Gabor-element was equal to the background luminance (49 cd/m 2 ). By selecting two orientations of the Gabor-elements amongst four possible orientations (0, 45, 90 and 135 deg) we obtained, for each pattern, two texture sub-regions separated by a texture-edge (Fig. 1). The texture-edge was located either between the two extreme stripes of elements (the up/down rows or the left/right columns) or between the two central stripes. In the first case, we will refer to the area with the larger number of iso-oriented elements as the larger texture-region. Segmented textures were differentiated on the basis of two distinct configurational properties (Fig. 1): - collinearity at the edge: the texture-elements (texels) in one of the two stripes abutting the edge, were either iso-oriented and collinear to each other, (in this case parallel to the edge), or iso-oriented and non-collinear (in this case non-parallel to the edge). - congruency in the larger texture-region: the texels in the larger textureregion were always iso-oriented and collinear to each other (except in Experiment 4) but their orientation was either congruent (parallel to the edge) or non-congruent (non-parallel to the edge) Subjects Subjects were aged years, all volunteers with normal or corrected-to-normal visual acuity. All the participants, except the authors, were ignorant of the purposes of the experiments. Each of the twelve subjects executed two experiments in random order: six participated in Experiments 1 and 2 and six in Experiments 3 and 4. Two new naïve subjects and the authors participated in Experiment Task Subjects performed a binary classification task and were asked to discriminate, by pressing one of two alternative keys, the orientation of the texture-edge (horizontal vs vertical).

3 E. Giora, C. Casco / Vision Research 47 (2007) Fig. 2. Events in a trial: first, a central fixation point was presented for 1000 ms on a grey background, followed (no interval) by a masking texture for 300 ms, and then (no interval) the test matrix, whose duration changed randomly in a block amongst five levels (20, 40, 60, 80, 100 ms); this was finally interrupted (no interval) by a second 300-ms mask. Fig. 1. Example stimuli. (a) Textures used in Experiment 1, with deg orientation contrast at the edge: elements in one side of the edge were either collinear (left) or non-collinear (right). (b) Textures used in Experiment 2, with an orientation contrast of 90 deg at the edge: texel overall-orientation in the larger region was either non-congruent (left) or congruent (right) with boundary orientation; in the symmetrical control stimulus (centre) the congruency distinction does not apply. (c) Textures used in Experiment 3: the conditions were as in Experiment 2 but with 45/ 135 deg orientation contrast at the edge. (d) Textures used in Experiment 4: non-congruent and congruent stimuli of Experiment 3 and the same patterns in jittered condition. (e) Textures used in Experiment 5: symmetrical pattern 0/90 deg of Experiment 2 and a new symmetrical pattern 45/135 deg Procedure The procedure is illustrated in Fig. 2. Each trial started with a central fixation point presented for 1000 ms on a grey background; the test matrix was preceded and followed (with no interval) by a 300 ms mask composed of randomly oriented Gabors. Test matrix durations were varied in five levels (20, 40, 60, 80 and 100 ms), randomly in a block. At the second mask offset, the screen turned black until pressing of a key by the subject to start a new trial. Each block consisted of a random presentation of eight repetitions of each orientation, duration level and configurational condition (collinear vs non-collinear in Experiments 1 and 5, in total 160 trials; congruent, control and non-congruent, in Experiments 2 and 3, in total 240 trials; congruent-collinear, congruent-jittered, non-congruent-collinear, noncongruent-jittered in Experiment 4, in total 320 trials). Before the experiment started, each subject performed a practice session in which visual feedback was given after each trial, depending on whether the response was correct or incorrect Data analysis Temporal thresholds, defined as the estimated duration for 75% accuracy, were calculated by Probit analysis (Finney, 1971). Repeated measures ANOVAs were used to compare individual thresholds. The sphericity of the data was tested with Mauchly s test (Howell, 2002). When the sphericity assumption was not supported by that test, Greenhouse Geisser correction was applied. Since the data for the two orientations were not statistically different, they were not treated separately. Post hoc comparisons were carried out using Bonferroni s correction. 3. Experiments and results 3.1. Experiment 1 It is well known that collinearity at the edge facilitates edge segmentation (Caputo & Casco, 1999; Casco et al., 2004, 2005; Wolfson & Landy, 1995). Experiment 1 was designed to confirm this facilitatory effect. In both stimuli, orientation contrast at the edge was the same (45 deg) and the texture field was filled with collinear elements oriented either at 45 or 135 deg, with equal probability. For both cases, elements in the extreme stripe were either collinear to each other and parallel to the edge (Fig. 1a, left) or non-collinear and non-parallel (Fig. 1a, right). Fig. 3 shows that accuracy increased with duration. Temporal thresholds were lower (31.9 vs 57.6) when texels in the furthest stripe were collinear with the edge [t (5) = 5.8, p <.005]. These results confirmed that segmentation is improved by collinearity at the edge Experiment 2 In Experiment 2, as in Experiment 1, elements in the furthest stripe were either collinear and parallel to the edge (Fig. 1b, left) or non-collinear and non-parallel (Fig. 1b, right). The collinear elements in the larger texture-region had 90 deg orientation contrast with those in the stripe. In this way, both stimuli had one or the other stripe abutting the edge having elements collinear and parallel to the edge. This should balance the collinearity effect of Experiment 1. However, the elements in the larger region were always collinear but either parallel (congruent) to the edge

4 882 E. Giora, C. Casco / Vision Research 47 (2007) Fig. 3. Results of Experiment 1. Mean accuracy of the six subjects is plotted as a function of stimulus duration (ms), separately for collinear (continuous line) and non-collinear (dotted line) patterns. (in Fig. 1b, right) or orthogonal (non-congruent) to the edge (Fig. 1b, left). Thus, only the congruency factor could account for a difference between the two conditions. As a control for these stimuli with a larger texture-region made up of seven stripes, a stimulus was used presenting identical orientation contrast at the edge, but with two sub-regions of identical size four stripes on each side of the edge, either parallel or orthogonal to the edge (Fig. 1b, centre). If edge saliency depended on collinearity at the edge only, we predicted no difference in the three conditions. On the other hand, if discrimination was based on region-based analysis, as in Caelli (1985) account, we expected better performance for the non-congruent condition and worse for the congruent. Results, reported in Fig. 4, showed that accuracy increased with duration. The effect of stimulus was significant [F (2,10) = 7.4; p <.05]. Post hoc comparisons showed that temporal thresholds were lower in the non-congruent (29.7 ms) than in the congruent (52.2 ms) condition [p <.05]. Temporal thresholds in the control condition were in between (32.5 ms), indicating the congruent condition to be more difficult than the control, where the two sub-regions segmented by the edge had the same size. Instead, the difference between control and non-congruent was weak. These results support the idea in agreement with region-based models that collinear elements in the larger texture improve segmentation when non-parallel to the edge. Note that this effect of non-congruency cannot be confused with collinear facilitation in the smaller region. Indeed, by decreasing the size of the larger non-congruent region in the control pattern, despite a larger number of collinear stripes parallel to the edge in the other sub-region, edge saliency was slightly reduced. Finally, the results cannot be accounted for by a difference in eccentricity that would predict a better performance for the control than for both patterns with edge located at a more external stripe of elements Experiment 3 Since in Experiments 1 and 2 orientation contrast differed, we replicated Experiment 2 to check the effect of orientation contrast. One of the stimuli used in Experiment 3 consisted of the same collinear stimulus used in Experiment 1, which was also non-congruent (Fig. 1c, left), and its symmetrical version, which was congruent (Fig. 1c, right). In both stimuli the edge was defined by 45 deg orientation contrast. Moreover, as in Experiment 2, in both stimuli elements were collinear on one side and non-collinear on the other; this should lead to an identical facilitation of collinearity at the edge. Data obtained with these two stimuli with a larger texture-region made up of seven stripes (as in Experiment 2) were compared with those in a control condition (Fig. 1c, centre) with no larger texture-region (four stripes on each side of the edge). The results, given in Fig. 5, showed an effect of duration similar to that obtained in Experiment 2. The ANOVA revealed a significant effect of stimulus [F (2,10) = 5.6; Fig. 4. Results of Experiment 2. Mean accuracy of the six subjects is plotted as a function of stimulus duration (ms), separately for congruent (squares), non-congruent (circles) and control (triangles) stimuli. Fig. 5. Results of Experiment 3. Mean accuracy of the six subjects is plotted as a function of stimulus duration (ms), separately for congruent (squares), non-congruent (circles) and control (triangles) stimuli.

5 E. Giora, C. Casco / Vision Research 47 (2007) p <.05]. Post hoc comparisons showed that temporal thresholds were lower in the non-congruent (26 ms) than in the congruent (55 ms) condition [p <.02]. Temporal thresholds in the control condition were in between (34 ms). With these data we confirm that the difference in saliency cannot be explained by collinearity at the edge. The non-congruency effect instead reflects a facilitation from the larger texture-region. The results of Experiments 1, 2 and 3 suggest two distinct contextual and non-local effects. One results from elements along the edge being collinear and parallel. The other results from elements in the larger texture-region that, if collinear and non-parallel to the edge, facilitate segmentation an effect that increases with the elements being linked (i.e. with texture size becoming larger). This second effect is consistent with a region-based account, predicting that collinear elements in the larger texture facilitate detection when non-parallel to the edge. Note that this explanation would predict that perturbation of collinearity should impair edge segmentation. However, also a local inhibitory effect of individual remote texels oriented as the edge in the congruent condition could account for the reduced saliency of the edge. Indeed, close facilitatory elements are present in all the three conditions but remote, inhibitory (Wolfson & Landy, 1999) texels are more numerous in the congruent, less numerous in the control, and absent in the non-congruent. Consequently, for this local explanation the crucial factor would be the individual orientation of texels, independently of their groupings by collinearity. To test this possibility we repeated Experiment 3 by presenting jittered and non-jittered patterns with equal probability. The local account would predict a similar performance Experiment 4 In Experiment 4, we jittered the spatial position of each element by 10 arcmin, in a random direction, to impede element collinearity (Fig. 1d). If facilitation in edge segmentation resulted from grouping of collinear elements both along the edge and in the larger texture, we predicted a stronger effect of jittering in the non-congruent condition than in the congruent condition. Fig. 6 shows that accuracy increased with duration. The ANOVA revealed a significant effect of congruency [F (1,5) = 30.1; p <.005] and of the jittering congruency interaction [F (1,5) = 7.0; p <.05]. Post hoc showed that the effect of jittering was significant only in the non-congruent condition [p <.05]. Oddly, jittering seems to improve performance in the congruent condition, but the effect is consistent only at 60 ms. The results for the non-congruent condition rule out the possibility that the facilitation comes from local inhibitory effect, because of remote individual texels iso-oriented with the edge, independent of grouping by collinearity. The data instead suggest that the non-congruency effect is due to region analysis, based on contextual facilitation resulting Fig. 6. Results of Experiment 4. Mean accuracy of the six subjects is plotted as a function of stimulus duration (ms), separately for jittered (dotted lines) and non-jittered stimulus (continuous lines), both congruent (squares) and non-congruent (circles). from excitatory lateral interactions amongst co-axial filters (Lamme & Roelfsema, 2000) Experiment 5 Overall, the results of previous experiments suggest that the relationship between the orientation of stripes resulting from collinear grouping and boundary orientation is the crucial factor. Nonetheless, one can argue that the enhanced discrimination for the non-congruent condition could result from having collinearity facilitation in both regions, leading to a strengthening and linking of the 1ststage filter output. Therefore, a simple edge-based mechanism would predict an increment of gradient detector response following a stronger activation due to collinearity. Consequently, the final response of edge extraction would be higher in the non-congruent condition, where collinearity is present in both the larger region and the more external stripe. Another possibility is that the interaction between stripes and edge orientation could actually be important, simply because of an inhibitory effect from remote stripes parallel to the edge. To distinguish amongst alternative explanations, we compared the control condition of Experiment 2 with stripes in the texture-region, both parallel and orthogonal to the edge (symmetrical 0/90 deg, see Fig. 1e, left) with a pattern having texels in both regions collinear, but neither parallel nor orthogonal to the boundary (symmetrical 45/ 135 deg, see Fig. 1e, right). If facilitation in the non-congruent condition resulted simply from collinearity on both sides of the edge, one would not expect the performance to differ. Furthermore, the absence of inhibition from remote stripes parallel to the edge should enhance edge saliency in the symmetrical 45/135 deg. Only, if the higher saliency of the non-congruent stimulus resulted from both facilitation at the edge and facilitation by region-based analysis, would we expect better performance in the symmetrical 0/90 deg condition. Results are shown in Fig. 7. Data appear coherent with those already obtained by Wolfson & Landy (1995) with

6 884 E. Giora, C. Casco / Vision Research 47 (2007) Fig. 7. Results of Experiment 5. Mean accuracy of the four subjects is plotted as a function of stimulus duration, separately for the symmetrical 0/90 deg (continuous lines) and 45/135 deg (dotted line) patterns. similar patterns. Temporal thresholds are much lower in the symmetrical 0/90 deg (31.5 ms) condition than in the symmetrical 45/135 deg (68 ms) condition [t (3) = 6.62, p <.01]. These results rule out the possibility that a facilitation depended on the presence of collinearity in both regions. Moreover, the data do not indicate an inhibitory effect resulting from parallel stripes remote from the edge. Instead, the important variable is the relationship between the orientation of stripes resulting from collinear grouping and the orientation of the boundary. 4. Discussion In summary, the five experiments showed that segmentation was affected by configurational factors both at the edge and from the texture-region. Experiment 1 showed higher saliency for edges with abutting texels on one side iso-oriented and parallel to it. Experiment 2 showed a new facilitatory effect resulting when collinear elements in the larger texture are non-parallel to the edge. This result, confirmed by Experiment 3, is compatible with a regionbased mechanism. The finding that jittering reduces the facilitation from the region (Experiment 4) rules out explanations based on local effects of individual texels, independent of grouping by collinearity (Wolfson & Landy, 1999). Experiment 5, showing better performance in the symmetrical 0/90 deg than in the symmetrical 45/135 deg, indicates that facilitation results not just from a simple effect of collinearity per se in both regions, but because of the specific relationship between the orientation of stripes resulting from collinear grouping and the orientation of the boundary. It has also been shown that remote parallel stripes do not have an inhibitory effect on edge extraction. Overall, the main finding is that texture segmentation is facilitated not only when abutting elements are parallel to the edge but also from groupings in the larger texture, non-parallel to edge orientation. The two facilitatory effects cannot be interpreted within the framework of energy models based on Filter-Rectify- Filter operations. The underlying assumption of these models is that the detection of texture differences does not occur locally (Nothdurft, 1992) but results from a global operation that is based on relative (not absolute) orientation at the level of local 1st-order and global 2nd-order filtering. The dependency on global orientation contrast, predicted by Filter-Rectify-Filter models, does not, however, manage to explain the new finding of a facilitation from the texture-region, because these models should be indifferent to whether texture is larger or smaller. Therefore, our effects are necessarily based on the facilitatory interaction between spatial channels and a model explicitly based on these interactions has to be considered to account for them. The idea that spatial channels, rather than being independent, spatially interact, has often been contemplated by models of texture perception. For example, Caelli s (1985) model proposed linking operations that follow the Gestalt rules for grouping e.g., proximity, similarity and good continuation (Beck, Prazdny, & Rosenfeld, 1983) based on spread of neural activity emanating from highly stimulated detectors. It is known from the pioneering work of Gestalt psychology (Wertheimer, 1922) that iso-oriented and collinear elements group together, and recent investigations demonstrate that lateral interactions in the primary visual cortex can account for grouping phenomena (Gilbert, Ito, Kapadia, & Westheimer, 2000; Stettler, Das, Bennett, & Gilbert, 2002; Zipser, Lamme, & Schiller, 1996). These recent neurophysiological data lead to the assumption that regionbased mechanisms treat neighbouring texture-regions as belonging to the same texture if texels are both iso-oriented and collinear (i.e. texels and receptive field orientation are co-axial). A region-based mechanism could account for edge segmentation at a 1st-stage of processing with no need for energy-gradient extraction, explaining the non-congruent facilitation. On the other hand, that mechanism would fail to detect the edge in the congruent condition, because spreading of the neural activation when texels are iso-oriented and collinear would occur in the same direction for all stripes and produce, in the stripe with no collinear elements, less salient groups. Because of this failure of the region-based mechanism, activation of an edge-based mechanism is required, at a successive stage of processing, to account for edge segmentation in the congruent condition. Note that this two-stages explanation is based on the strong assumption that region-based analysis precedes edge extraction. Further data suggest that grouping by collinearity occurs before edge extraction (Giora & Casco, 2006). Lateral connectivity occurring in V1 could well be the underlying neural mechanism (Stettler et al., 2002). However, there is another possible interpretation for the facilitation at the edge and from the larger texture. At the edge, grouping between elements parallel to the edge can modulate the output of the edge-based mechanisms at some stage during the Filter-Rectify-Filter operations, probably at the level of the 1st-stage filtering or after rectification. Instead, facilitation from the region could reflect

7 E. Giora, C. Casco / Vision Research 47 (2007) activation of 2nd-order filters selective for orientation, spatial-frequency and length: facilitation could result when 2nd-order filters are stimulated by elongated groups of texels, formed by collinear linking with the same spatialfrequency and length but different orientation on either side of the edge. Several sets of data suggest that 2nd-order channels are orientation and spatial-frequency tuned (Arsenault, Wilkinson, & Kingdom, 1999; Kingdom & Keeble, 1996; Kwan & Regan, 1998). Landy & Oruç (2002) proposed that the 2nd-stage filters receive differently-oriented 1st-stage inputs to the centre and surround of the receptive field. They are therefore tuned to orientation differences rather than orientation per se. If these 2nd-order mechanisms were also end-stopped (Yu & Levi, 1997) they would prefer stimuli on the two sides of their receptive field having different orientation but identical spatial-frequency and length. Assuming groupings of iso-oriented and collinear elements before 2nd-order filtering, their optimal activation would occur when groups of appropriate orientation contrast (matching the orientation contrast preferred by a filter) fell on the receptive field. This hypothesis of 2ndorder filters selective not only for spatial-frequency and orientation contrast but also for length, would explain why congruent is worse than control and control is worse than non-congruent. Indeed, only in the non-congruent condition do oriented groups have the same orientation and length as the edge, whereas in the control and congruent conditions the length of groups is shorter in one side of the edge. To conclude, our account contrasts with the prominent view that segmentation is a first stage in the hierarchical organization of visual processing modules of increasing complexity edges, surfaces, objects (Marr, 1982). Our results indicate that overall-orientation information biases texture segmentation, thereby supporting the suggestion (Oliva & Torralba, 2001) that the holistic representation of an artificial image in the laboratory, as well as a scene in the real world, is an operation taking place earlier than the segmentation itself. Acknowledgments The work described in this paper was supported by a Grant (PRIN 2003) for the Ph.D. programme of the first author. We would like to thank Dr. G. Campana, Prof. M.S. Landy and Prof. M.J. Morgan for a critical reading of the manuscript and their useful suggestions. References Ariely, D. (2001). Seeing sets: representation by statistical properties. Psychological Science, 12, Arsenault, A. S., Wilkinson, F., & Kingdom, F. A. A. (1999). Modulation frequency and orientation tuning of second-order texture mechanisms. Journal of the Optical Society of America A, 16, Beck, J., Prazdny, K., & Rosenfeld, A. (1983). A theory of textural segmentation. In J. Beck, B. Hope, & A. Rosenfeld (Eds.), Human and machine vision (pp. 1 38). New York: Academic. Ben Shahar, O. (2006). Visual saliency and texture segregation without feature gradient. Proceedings of the National Academy of the Sciences of the United States of America, 103, Caelli, T. (1985). Three processing characteristics of visual texture segmentation. Spatial Vision, 1, Caputo, G. (1996). The role of the background: texture segmentation and figure-ground segmentation. Vision Research, 36, Caputo, G., & Casco, C. (1999). A visual evoked potential correlate of global figure-ground segmentation. Vision Research, 39, Casco, C., Campana, G., Grieco, A., & Fuggetta, G. (2004). Perceptual learning modulates electrophysiological and psychophysical response to visual texture segmentation in humans. Neuroscience Letters, 371, Casco, C., Grieco, A., Campana, G., Corvino, M. P., & Caputo, G. (2005). Attention modulates psychophysical and electrophysiological response to visual texture segmentation in humans. Vision Research, 45, Field, D. J., Hayes, A., & Hess, R. F. (1993). Contour integration by the human visual system: evidence for a local association field. Vision Research, 33, Finney, D. J. (1971). Probit analysis. London: Cambridge University Press. Gilbert, C., Ito, M., Kapadia, M., & Westheimer, G. (2000). Interactions between attention, context and learning in primary visual cortex. Vision Research, 40, Giora, E., & Casco, C. (2006). Seeing texture figures before boundaries. Perception, 35(Suppl.), 74. Hess, R. F., & Dakin, S. C. (1997). Absence of contour linking in peripheral vision. Nature, 390, Howell, D. C. (2002). Statistical methods for psychology (5th ed). Belmont, CA: Duxberry Press. Kingdom, F. A. A., & Keeble, D. R. T. (1996). A linear systems approach to the detection of both abrupt and smooth spatial variations in orientation-defined textures. Vision Research, 36, Kwan, L., & Regan, D. (1998). Orientation-tuned spatial filters for texture-defined form. Vision Research, 38, Lamme, V. A. F. (2004). Beyond the classical receptive field: contextual modulation of V1 responses. In L. M. Chalupa & J. S. Werner (Eds.). The visual neurosciences (Vol. I, pp ). Cambridge, MA: MIT Press. Lamme, V. A. F., & Roelfsema, P. R. (2000). The distinct modes of vision offered by feedforward and recurrent processing. Trends in Neurosciences, 23, Landy, M. S., & Graham, N. (2004). Visual perception of texture. In L. M. Chalupa & J. S. Werner (Eds.). The visual neurosciences (Vol. II, pp ). Cambridge, MA: MIT Press. Landy, M. S., & Oruç, I. _ (2002). Properties of second-order spatial frequency channels. Vision Research, 42, Lee, T. S. (1995). A Bayesan framework for understanding texture segmentation in the primary visual cortex. Vision Research, 35, Li, W., & Gilbert, C. D. (2002). Global contour saliency and local colinear interactions. Journal of Neurophysiology, 88, Malik, J., & Perona, P. (1990). Preattentive texture discrimination with early vision mechanisms. Journal of Optical Society of America A, 7, Marr, D. (1982). Vision. San Francisco: Calif.: Freeman. Nothdurft, H. C. (1992). Feature analysis and the role of similarity in preattentive vision. Perception & Psychophysics, 52, Oliva, A., & Torralba, A. (2001). Modeling the shape of the scene: a holistic representation of the spatial envelope. International Journal of Computer Vision, 42, Olson, R. K., & Attneave, F. (1970). What variables produce similarity grouping? American Journal of Psychology, 83, 1 21.

8 886 E. Giora, C. Casco / Vision Research 47 (2007) Parkes, L., Lund, J., Angelucci, A., Solomon, J. A., & Morgan, M. (2001). Compulsory averaging of crowded orientation signals in human vision. Nature Neuroscience, 4, Polat, U., & Sagi, D. (1993). Lateral interactions between spatial channels: suppression and facilitation revealed by lateral masking experiments. Vision Research, 33, Polat, U., & Sagi, D. (1994). The architecture of perceptual spatial interactions. Vision Research, 34, Roelfsema, P. R., Lamme, V. A., Spekreijse, H., & Bosch, H. (2002). Figure-ground segregation in a recurrent network architecture. Journal of Cognitive Neuroscience, 14, Sagi, D. (1991). Spatial filters in texture segmentation tasks. In B. Blum (Ed.), Channels in the visual nervous system: neurophysiology, psychophysics and models (pp ). London and Tel Aviv: Freund Publishing House. Stettler, D. D., Das, A., Bennett, J., & Gilbert, C. D. (2002). Lateral connectivity and contextual interactions in macaque primary visual cortex. Neuron, 36, Yu, C., & Levi, D. M. (1997). End stopping and length tuning in psychophysical spatial filters. Journal of the Optical Society of America A, 14, Wertheimer, M. (1922). Untersuchungen zur Lehre von der Gestalt I. Psychologische Forschung, 1, Wolfson, S. S., & Landy, M. S. (1995). Discrimination of orientationdefined texture edges. Vision Research, 35, Wolfson, S. S., & Landy, M. S. (1999). Long range interactions between oriented texture elements. Vision Research, 39, Zipser, K., Lamme, V. A. F., & Schiller, P. H. (1996). Contextual modulation in primary visual cortex. Journal of Neuroscience, 16,