Object Localization in Blurred and Jumbled Scenes: Differences between Myopic and Emmetropic Observers

Abstract

purpose. To compare the effects of transient blur constraints on the perception of natural scenes in myopic and emmetropic observers. When natural images are blurred, the global structure of the scene—its context—becomes essential for decision making. The authors also aimed to control whether the difference in performance between myopes and emmetropes, if any, resulted from the ability to use contextual information or the ability to process blurred features.

methods. Twenty-four subjects (13 emmetropes, 11 myopes) participated in the study. Subjects were instructed to fixate a target in a scene. Low-pass–filtered images were briefly displayed (100 ms) with normal or jumbled structures. The experiment was repeated three times to assess the ability of subjects to adapt to the imposed blur constraints.

results. Results showed that myopes performed better than emmetropes only when scene context was available. Myopes also adapted faster to the jumbling constraint. However, after three repetitions, both groups reached the same level of performance.

conclusions. The present study demonstrated that myopes did not perform better than emmetropes in a transiently blurred environment, but they exhibited better and faster ability to adapt their perceptual strategies to the visual constraints by learning to make use of the relevant information available in the images displayed.

Amajor advantage of the human visual system, compared with artificial systems, is its flexibility and its apparent ability to adapt its processing strategies according to prevailing perceptual conditions. Numerous reviews demonstrate these adaptation capabilities. For example, recent studies ^{1 2} have shown that when observers are exposed to prolonged blurred stimuli, their performance is progressively better on tests of visual acuity and contrast sensitivity. There is, however, large intersubject variability. Refractive errors may constitute a factor in such variability. Indeed, numerous studies have shown that myopes are more efficient than emmetropes in blurred visual environments. ^{3 4 5 6 7}  

Despite the differences in procedures, previous experimental studies share two common points: they all examined the effect of sustained blur on visual perception and they all used as stimuli simple forms (e.g., letters) displayed on uniform backgrounds. The experiment reported here aimed to assess the visual behavior of myopes and emmetropes on perceptually mediated tasks under a range of viewing conditions. In particular, the experiment tested the effect of transient blur on the perception of natural objects displayed in their usual environment. It was hypothesized that myopes would perform better than emmetropes because they are more efficient at analyzing blurred visual information. 

However, because we used natural images, we had to consider the role of the scene context on the subject’s performance. Indeed, it has been suggested that context may help to localize targets. ^{8 9} What do we mean by context? Contextual information, as defined here, involves the “logical” spatial relationships existing between various elements in a scene. These logical relationships, derived from perceptual experience, depend on our previous knowledge of the visual world and on the various rules governing proper spatial organization. Contextual information is mainly provided by the low spatial frequencies of the scene. ¹⁰ For images with high resolution, previous works ^{8 9} have shown that subjects could localize objects displayed in their natural environment without the involvement of any contextual information. Scene context remained merely optional, as long as the object’s prominence in its environment was sufficient; in such a case, the decision was only influenced by the analysis of the object’s individual features. However, the results also showed that when images were blurred or when the distinctive features of the object were poorly recognizable, scene context became useful to localize the object. ^{8 9} Therefore, the better ability of myopes to localize targets in blurred images could be attributed to better ability to use contextual information or process blurred object features. In earlier work, we controlled the intervention of scene context with the use of jumbled images. The jumbling technique ¹¹ required the picture to be broken down into six parts that were then jumbled to disturb the spatial relationships between various scene elements. In jumbled scenes, context was no longer available; the observer had to rely on local object features to achieve the task. 

We hypothesized that a group of myopic subjects would perform better than a group of emmetropes at localizing targets when images were low-pass filtered. If this difference resulted from the way both groups used context, it would appear only for images not jumbled. Emmetropes and myopes would exhibit the same level of localization for jumbled images. On the other hand, if the difference was related to the better ability of myopes to process and interpret the blurred local features of the targets, it would remain constant whatever the global structure of the scene. 

Methods

Subjects

The experiment involved 24 subjects (12 women, 12 men; age range, 25–40 years; mean age [±1 SD], 30.29 ± 5.39 years). Two groups were distinguished according to the subject’s refractive error: 13 subjects were considered emmetropes (plano to +0.75 D; mean refractive error, +0.14 ± 0.23 D, with astigmatism <1.00 D); 11 subjects comprised the group of myopes (–0.75 D to –2.50 D; mean refractive error, –1.32 ± 0.63 D, with astigmatism <1.00 D). For the experimental conditions of this study, subjects wore their current spectacles and had visual acuity of at least 10/10 (only one subject did not wear refractive correction full time). Subjects’ sensitivity to contrast was tested (Visteck contrast sensitivity chart), and all were found to be normal over the seven spatial frequencies considered (0.75–18 cpd). Subjects were volunteers from the laboratory. The study adhered to the tenets of the Declaration of Helsinki. Informed consent was obtained from each subject after explanation of the general purpose of the experiment (i.e., the ability of persons with different refractive errors to localize objects against various types of background). They received no payment for participation. 

Setup

Test images were displayed to subjects on a 21-inch color monitor (Trinitron; Sony, Tokyo, Japan). Each subject sat 60 cm from the screen, head resting on a chin rest. To ensure that the subjects were not disturbed by any reflection on the screen, the test was conducted in a room with no light source other than the monitor. 

Stimuli

Visual stimuli used were grayscale pictures from the Corel database (Corel Corporation, Ottawa, Ontario, Canada) and represented natural scenes from everyday life (country, city, or home scenes). Six pictures were selected and were chosen according to size, location, and environment of objects. The target was located in the central block of a grid, made up of five horizontal × three vertical blocks (Fig. 1) . Objects were evenly distributed throughout the entire surface of the picture. Each “original image” had an average luminance of 35 cd/m² and contained 2135 × 1440 pixels. On the virtual 15-block matrix, a selection mask corresponding to six blocks (3 × 2) was used to extract six smaller images of 1280 × 960 pixels, making up the initial image base displayed on the screen. The target had different positions on the screen in each of these smaller images. It could appear at any of the six possible locations. 

Three cutoff spatial frequencies (CFs), achieved using Gaussian filters, ensured a variation in the level of blur: 3, 1.5, and 1 cpd. Therefore, each filtering level (respectively identified as CF3.0, CF1.5, CF1.0; NF indicates nonfiltered image) corresponded to an increasing truncation of the frequency spectrum of images along high spatial frequencies. 

Context content of images was modified by cutting each image into six blocks (3 × 2) that were then randomly jumbled. ¹¹ This jumbling broke all coherent spatial relationships between various scene elements in the image (Fig. 2) . Coherent contextual information, available in the normal picture, was no longer present in the associated jumbled image. To compare the performance obtained for these two kinds of images, the target was systematically located in the same place in the normal and in the jumbled image. Furthermore, all blocks neighboring the target’s block were systematically moved in the jumbled image. 

All experimental conditions were combined for each image, generating 48 configurations (four spatial frequency filtering levels × two context levels × six positions). The six original images thus resulted in 288 different images or trials. 

Procedure

The experiment was broken down into three series. Each subject viewed, in random order, the 288 images during each series (S1, S2,S3); the same 288 images were seen three times. The duration of each series was approximately 10 minutes. Subjects performed only one series per day over three consecutive days. Before data were recorded, a preliminary training session helped the subject to become accustomed to the experimental setup and to the task, which was to localize the cued target. 

During each trial, the name of the object was first displayed for 500 ms as a text message. The test image was then displayed for 100 ms. After viewing, a matrix came on the screen displaying the six blocks of the image. The instruction was to click with the computer mouse on the correct location of the target. Feedback of the correct response was given, and then the next trial was activated (Fig. 3) . 

Analysis

Four independent variables were considered in this experiment: three within-subject variables (spatial frequency content, context structure, and series) and one between-subject variable (emmetropes versus myopes). Analysis of results included the correct response rate (i.e., dependent variable) expressed as a percentage of the number of trials. 

Results

Analyses first focused on the first experimental series. Three-way repeated-measures ANOVA was conducted with spatial frequency (SF) content and context as within-subject variables and refractive error as the between-subject variable. Our results showed that myopes were significantly better at localizing targets than emmetropes (refractive error, F[1;22] = 6.11; P = 0.02167; Fig. 4A ). All subjects made more errors when the spatial arrangement of scene elements was disturbed, or jumbled (context, F[1;22] = 189.43; P < 0.0001; Fig. 4B ). However, myopes were proportionally more disturbed by jumbling than emmetropes—that is, their performances with normal images were significantly more degraded by jumbling than were those of emmetropes (refractive error × context, F[1;22] = 9.01; P = 0.00657). Overall correct responses rate gradually decreased as spatial frequencies were deleted from the spectrum by filtering (SF content, F[3;66] = 110.41; P < 0.0001). Whatever the filtering level, performances systematically decreased as images were jumbled (context × SF content, F[3;66] = 0.68; P = NS; Fig. 5A ). Three-way interaction was close to significant (refractive error × context × SF content, F[3;66] = 2.46; P = 0.07029). Considering only the normal images (Fig. 5B) , our results showed that as the level of blur increased, myopes maintained better performance than emmetropes (refractive error, F[1;22] = 9.09; P = 0.00637; refractive error × SF content, F[3;66] = 3.58; P = 0.01839). However, when the images were jumbled (Fig. 5C) , no statistical difference was found between the groups, whatever the filtering level (refractive error, F[1;22] = 0.65; P = NS; refractive error × SF content, F[3;66] = 0.39; P= NS). 

The aim of the present study was to determine whether myopes and emmetropes were differently affected by blur. Results for normal images showed that the difference was significant and mainly attributed to the two higher levels of filtering (FC1.5 and FC1.0; see Fig. 5B ). Hence, the change in localization rate as a function of the series was assessed only for both higher levels of filtering. Considering the three series, the change in performance differed between myopes and emmetropes, whatever the contextual information available (refractive error × series, F[2;44] = 5.40; P = 0.00802 for normal images; Fig. 6A ; F[2;44] = 3.25; P = 0.04857 for jumbled images; Fig. 6B ). Considering the differences in performance between the first and the second series, our results showed that the increased localization rate in normal images was not statistically different for both groups (refractive error × series, F[1;22] = 1.83; P = NS). However, this difference was significant for jumbled images (refractive error × series, F[1;22] = 4.46; P = 0.04636). Between the second and the third series, the performance increased more for emmetropes than for myopes, whatever the global structure of the scene (refractive error × series, F[1;22] = 6.25; P = 0.02038 for normal images; F[1;22] = 5.15; P = 0.03338 for jumbled images). 

Discussion

The results of the present study describe a new feature of the visual perception of myopes. Previous works has provided evidence that myopes exhibit better ability than emmetropes to adapt to sustained optical blur. ^{3 4 5 6 7} In the present study, when blurred images were briefly displayed, the results confirmed that myopes managed blur constraints more easily than emmetropes. Indeed, regarding the results of normal images, correct response rates from myopic observers were significantly less affected by increasing blur than from emmetropes. However, myopes and emmetropes performed the task with the same efficiency when scene context information was removed by jumbling. These results suggest that the better performance of myopes to localize targets in briefly displayed blurred images is due to better ability to use contextual information rather than to better ability to process blurred object features. 

Furthermore, the present results show that when exposure to the visual constraints (blurred or jumbled images, or both) was repeated, the improvement in performance of myopes and emmetropes was markedly different. During the first series, regarding both higher filtering levels, the better efficiency of myopes is linked to their ability to find the correct location of the target because of the scene context. From the first to the second series, performance improvement for both groups was similar for normal images. However, for jumbled images, the correct response rate increased more for myopes than for emmetropes. How could this difference be interpreted? In a comparable framework, a previous study showed that the target localization rate for jumbled and blurred images could be enhanced through an adaptation process with repeated exposure to jumbled, blurred images. ⁹ The term adaptation may have various meanings, depending on the research domain. In the present work, we define the adaptation phenomenon as a change of behavior in reaction to or induced by modifications of the usual perceptual conditions. This definition is also consistent with studies on adaptation to sustained dioptric blur in which the change of behavior is classically expressed in terms of changes in visual acuity or perceptual sensitivity. Results of our previous experiments ^{8 9} reflect an adaptation phenomenon by which subjects learned to extract the local relevant features that allowed them to distinguish the target from its direct environment and from the other objects of the scene. These local relevant or distinctive features were called the target spatial signature. When observers succeeded in extracting the spatial signature of each target, they increased their ability to localize them in jumbled scenes, relying more on the relevant and invariant local information than the global fluctuating context. The present results suggest that, during the second series, the adaptation process (learning the target spatial signatures) would be more efficient for myopes than for emmetropes, allowing the former to increase more rapidly their object localization performance for jumbled images. Results of the last series (S3) showed that myopes were not better than emmetropes at performing the target localization task in blurred and jumbled or normal scenes. When enough time was given to emmetropes (enough repetition of the visual constraint) to efficiently use the context information and to extract target spatial signatures, they reached the same level of performance as myopes. 

The scientific literature provides few explanations of what happens during the adaptation process to sustained optical blur. Rosenfield et al. ⁷ show that the blur adaptation effect noticed for myopic observers is not linked to a modification of the refractive error. These authors ⁷ and others ^{1 6} suggest that increasing visual acuity after sustained optical blur was a cortical rather than a retinal phenomenon. Mon-Williams et al. ¹ describe their adaptation process as a neural adjustment induced by the prolonged exposure to visual environments involving only low spatial frequency channels. Basically, they suggest that myopes’ ability to adapt to prolonged blur is due to their history of inexact refractive correction. ¹ They might have learned to interpret their blur. However, little is known about this learning process. 

Initially, our study seemed to be far from this topic. Indeed, blurred stimuli were briefly displayed. The blur was obtained through low-pass filters different from what myopes experience by removing their optical correction. We used natural scenes and asked the subjects to perform a more normal task than the visual acuity tasks classically considered in the literature. However, the better ability of myopes to adapt to our constraints and to sustained optical blur could be linked to a similar learning effect, what has been called a perceptual learning effect. The notion of perceptual learning refers to relatively long-lasting changes in the perceptual system to improve the ability to respond to unusual external constraints induced by the environment. ¹² Considering the low refractive error of the myopic group, the fact that they were all fully corrected, and the properties of our filters, we assumed that none of the subjects involved in the present study had ever experienced such visual constraints (i.e., strong low-pass filtering and jumbling). We aimed to study the properties of their reactions to this unfamiliar visual environment. Two levels of reaction were shown: when subjects were exposed to transient blur constraint for the first time, they learned to rely more on the context of the scene to localize the targets. Then, with repeated exposure, they learned to catch locally the targets’ distinctive features, becoming free of the fluctuating global context. Even if the present experimental procedure did not correspond to the procedures specifically designed to study the perceptual learning process, ¹³ the results agreed with the general definition of this phenomenon and may be interpreted in terms of perceptual learning. 

The adaptation to sustained blur could also be due in part to a perceptual learning effect. Indeed, the visual system of low myopes and, more generally, of fully corrected myopes is often exposed to well-defined rather than blurred visual information. When they remove their correction, they experience an unfamiliar situation, as in our experiment. We suggest that the recovery of better visual acuity results from a learning mechanism: myopes learn to use in the available signal the relevant information—previously unknown or unnecessary because of the high resolution of the stimulations usually conveyed by their visual system—for decision making in the given visual task. For instance, in visual acuity tasks, they could learn to interpret the global structure of the letters rather than to rely on their detailed features. Myopes would be more efficient than emmetropes in this learning process, but our results suggest that given more time emmetropes could also learn to manage such visual constraint. With perceptual learning as only one part of optical blur adaptation (the memory of past blurred vision experiences would be another component of adaptation), emmetropes would require long exposure to a blurred world before they could obtain the same level of improvement in visual acuity that myopes have. Further investigation is needed to determine the potential contribution of perceptual learning in the process of sustained optical blur adaptation. 

The adaptation processes observed in our study, and perhaps in previous works on prolonged exposure to optical blur, might have resulted, at least in part, from changes in the visual strategies induced by a perceptual learning mechanism. Oliva and Schyns ¹⁴ clearly show this flexibility of perceptual strategies: the processing schema implemented by the visual system to meet a certain task can be greatly influenced and redirected by a preliminary sensitization period. However, they also demonstrated that even if the visual system was sensitized to a single kind of relevant information used in the image to perform the task, nonrelevant information was nonetheless integrated, even though it played no part in decision making. Oliva and Schyns ¹⁴ argue in favor of an active selection process. The hypothesis we developed from the results obtained in the present experiment follows the same lines. Adaptation to our visual constraints (blur and jumbling) would be linked to active selection in the image of the relevant information, previously unknown or unnecessary because of the usual broad spatial frequency spectrum available in images, allowing the visual system to optimize its performance. This phenomenon may be related to a perceptual learning effect. Adaptation to blurred scenes requires the visual system to learn to use the contextual information, and adaptation to blurred and jumbled scenes implies learning how to extract spatial signatures of the target. Our results show that myopes are more efficient than emmetropes in both learning processes. 

 Submitted for publication April 5, 2005; revised July 25 and October 17, 2005, and February 8, 2006; accepted June 23, 2006.

 Disclosure: G. Giraudet, None; L. Azavant, None

 The publication costs of this article were defrayed in part by page charge payment. This article must therefore be marked “advertisement” in accordance with 18 U.S.C. §1734 solely to indicate this fact.

 

Figure 1.

 

Method used to generate six smaller images from the original image. The original image is defined by 2135 × 1440 pixels. On the virtual 15 blocks matrix, a selection mask corresponding to six blocks (3 × 2) is used to extract six smaller images of 1280 × 960 pixels. With this arrangement, the target can be located in six different places, but semantic information remains relatively steady.

Figure 2.

 

Example of image jumbling. This procedure, borrowed from Biederman, ¹¹ is used to modify the contextual content of the scene by cutting it into six blocks (3 × 2) that are then randomly jumbled. The subject has to localize the horse sculpture. When images are jumbled, targets keep the same location.

Figure 3.

 

Sequence of events during a test. The sequence starts with the name of the target. The image is then displayed for 100 ms, after which the subject must indicate with the computer mouse the block in which the object is located. Then feedback is given on the correct target location.

Figure 4.

 

Correct response rates according to observers’ refractive error (A) and separately for normal and jumbled (B) images. Error bars represent 95% confidence interval. Asterisk indicates significant difference found by Scheffé post hoc test.

Figure 5.

 

(A) Correct responses rates for normal and jumbled images according to the spatial frequency content of the test images. NF, CF3.0, CF1.5, and CF1.0 correspond, respectively, to nonfiltered images and low-pass filtered images with cutoff frequencies at 3.0, 1.5, and 1.0 cpd. Myopes’ and emmetropes’ correct responses rates for normal (B) and jumbled (C) images according to the spatial frequency content of the test images. Error bars represent 95% confidence interval. Asterisk indicates a significant difference found by Scheffé post hoc test.

Figure 6.

 

Myopes’ and emmetropes’ correct response rates for normal (A) and jumbled (B) images according to the experimental series. S1 refers to the first time the 288 tests images are seen by observers. The same 288 images are seen in the second (S2) and third (S3) series. Error bars represent 95% confidence interval.