Simulation of Artificial Vision, III: Do the Spatial or Temporal Characteristics of Stimulus Pixelization Really Matter?

Abstract

purpose. In preceding studies, simulations of artificial vision were used to determine the basic parameters for visual prostheses to restore useful reading abilities. These simulations were based on a simplified procedure to reduce stimuli information content by preprocessing images with a block-averaging algorithm (square pixelization). In the present study, how such a simplified algorithm affects reading performance was examined.

methods. Five to six volunteers with normal vision were asked to read full pages of text with a 10° × 7° viewing window stabilized in central vision. In a first experiment, reading performance with off-line and real-time square pixelizations was compared at different resolutions. In a second experiment, off-line square pixelization was compared with off-line Gaussian pixelization with various degrees of overlap. In a third experiment, real-time square pixelization was compared with real-time Gaussian pixelization.

results. Results from the first experiment showed that real-time square pixelization required approximately 30% less information (pixels) than its off-line counterpart. Results from the second experiment, using off-line processing, revealed a restricted range of Gaussian widths for which performances were equivalent or significantly better than that obtained with square pixelization. The third experiment demonstrated, however, that reading performances were similar in both real-time pixelization conditions.

conclusions. This study reveals that real-time stimulus pixelization favors reading performance. Performance gains were moderate, however, and did not allow for a significant (e.g., twofold) reduction of the minimum resolution (400–500 pixels) needed to achieve useful reading abilities.

Currently, several research groups are working toward the development of visual prostheses for the blind. ^{1 2 3 4 5 6 7} Despite fundamental design differences (implantation site, image acquisition, and processing techniques), these approaches share common features that lead to several major constraints on the visual percepts that can be elicited. Envisioned devices consist of a finite number of discrete stimulation contacts, will be implanted at a fixed location in the eye, and will subtend only a fraction of the entire visual field. If one expects to restore useful vision to blind patients, these constraints have to be thoroughly considered. 

Our research group is part of a larger multidisciplinary research effort aiming to develop a subretinal implant. Our CMOS-Retina ^{8 9 10} is built to transform incident light on the retina into electric stimulation currents “in situ.” In this context, we have developed special experimental conditions (simulations) to explore the minimum requirements to restore useful artificial vision. 

Our simulations use low-resolution (pixelized) images that are projected in a “small” viewing area, stabilized at a fixed location in the visual field. We attempt to mimic the type of visual information provided by a retinal implant, using photodiode technology to transform incident light into an electric signal. With this methodological approach we explored, in a first study, ¹¹ the reading of isolated four-letter words. In central vision, accurate recognition was possible with pixelizations down to 286 pixels, distributed over a 10° × 3.5° viewing window. After a period of systematic training, comparable results were achieved with the same viewing window stabilized at 15° eccentricity in the lower visual field. In a second study, ¹² we explored full-page text reading under similar conditions. Tests were performed with a larger viewing window of 10° × 7° containing 572 pixels, that moved across the page of text under control of the subject’s eye movements. Performance was close to perfect with central vision. With eccentric vision, subjects achieved reading scores between 86% and 98% after a period of methodical training. 

In earlier studies, we used a simplified technique to simulate the limited number of stimulation contacts available in a visual prosthesis. Stimulus images were decomposed into a finite number of pixels with a simple block-averaging algorithm. This resulted in a mosaic of square pixels of various gray levels, the gray level within each pixel being constant (square pixelization). However, electrophysiological research ^{13 14 15} revealed that the patterns of neural activity elicited by electric stimulation of the retina depend on the strength of the stimulation current and that neural activation diminishes progressively with increasing electrode-to-neural target distance. These findings imply that phosphenes elicited by electrical stimulation of the retina should not be of constant luminosity and not of square shape. Furthermore, depending on the strength of the stimulation current, the percepts may develop from a collection of isolated phosphenes toward more continuous patterns with different degrees of overlap across neighboring phosphenes. 

One could argue that square pixelization is adequate to simulate the reduced information content of the stimuli transmitted by a retinal implant. In a given condition, the detailed shape of each pixel does not alter the overall information content of the image. However, studies on face recognition have demonstrated that detection is considerably hampered when images are decomposed into uniform square pixels. Harmon and Julesz ¹⁶ suggested that the oriented high-frequency noise introduced at block borders masks certain image features essential for recognition. Gestalt psychologists ^{17 18} further proposed that square pixelization distorts the image to the point of modifying its intrinsic gestalt properties. ¹⁹ Bachmann and Kahusk ²⁰ also suggest that the “block” constituents or pixels of the processed image compete for attention with the particular features of the image, thus affecting recognition. If one wants to avoid these drawbacks, square pixelization should be replaced by other types of image quantization featuring softer borders and allowing for variable amounts of overlap. 

Another shortcoming of our previous studies is that the pixelization algorithm was applied off-line over the entire original image (e.g., seven lines of full-page text). Subjects were allowed to scan this preprocessed image through a viewing window containing a subset of 572 pixels, the gray level of these “frozen” pixels being independent of the point of gaze on the image. This would not be the case in artificial vision systems, since stimulation intensity at each electrode contact would depend on the exact point of gaze relative to the image observed. For retinal implants transforming light falling on the retina into stimulation currents “in situ,” ^{4 7 10} this would happen due to eye movements. Head movements would act similarly in systems using an external head-mounted camera for stimulus generation. ^{1 2 3 5 6} In the case of reading, when focusing on a string of a few characters, its appearance would change on small eye (or camera) movements. Temporal cues seem to play a significant role in visual perception: the human visual system is optimized for detecting structural changes in dynamic images. A dynamic sequence of slightly different pixelized images may contain more information than one frozen pixelized image; therefore, dynamic (real-time) pixelization is likely to enhance information transmission to the visual system. Major object identification features (such as shape or location) are extracted from different spatial patterns (such as local contrast changes or relative position changes) resulting from image motion. Improved sensitivity for moving contrast changes, compared to their static equivalents, has previously been demonstrated. ²¹ Moreover, it has already been established that dynamic presentations lead to better performance in tasks like facial recognition. ^{22 23 24} Hence, if one wants more accurate simulations of artificial vision, pixelization should be performed in real-time and the intensity of each pixel should vary dynamically, according to gaze position. 

To our knowledge, psychophysical research using simulations of prosthetic vision has not been extensive so far. Reading and mobility were first studied by a group at the University of Utah. ^{25 26} Their head-mounted experimental setup consisted of a video camera sending images to a monochrome monitor that projected to the subject’s right eye (maximum viewing angle of 1.7°). Pixelization was achieved by overlaying the monitor with opaque masks containing a variable number of square perforations (pixels). Recently, another group at The Johns Hopkins University presented a series of experiments that used simulations specifically designed to mimic percepts evoked by retinal implants. ^{27 28 29} Different pixelization algorithms were used: a square pixelizing filter similar to the one presented in this article, a constant luminosity circular pixelizing filter, and a nonoverlapping Gaussian filter. Unfortunately, no direct comparison of the different pixelizing algorithms has been reported. Moreover, all these experiments neglected a fundamental aspect of artificial vision with a retinal implant: Viewing areas were not stabilized at fixed (eccentric) retinal positions. In more recent studies, the latter authors acknowledged that the stabilization of the viewing area on the retina can significantly affect performance (Dagnelie G, et al. IOVS 2004;45:ARVO E-Abstract 4223; Kelley AJ, et al. IOVS 2004;45:ARVO E-Abstract 5436), especially in visually demanding tasks such as reading. 

To validate our previous studies as well as to improve our simulation methods for future studies, we decided to investigate specifically the influence of the spatial and temporal characteristics of stimulus pixelization on reading performance. In the present study, we report a series of three paired comparisons of the effects of different pixelization methods on full-page reading. We compared reading performance: (1) between off-line square pixelization and real-time square pixelization of the image, (2) between off-line square pixelization and off-line Gaussian pixelization of the image, and (3) between real-time square pixelization and real-time Gaussian pixelization of the image. 

Methods

Subjects

Ten subjects aged between 23 and 41 years were recruited from the staff of the Geneva University Ophthalmology Clinic. All of them had perfect command of French, corrected visual acuity of 20/20 or better, and normal ophthalmic status. They were familiar with the purpose of the study and signed appropriate consent forms. All experiments were conducted according to the ethical recommendations of the Declaration of Helsinki and were approved by local ethics authorities. 

Experimental Setup

The stabilized projection of a 10° × 7° viewing window on the retina was achieved with a high-speed video-based eye and head-tracking system (EyeLink; SensoMotor Instruments GmbH, Berlin, Germany) and a high-refresh-rate monitor (Fig. 1) . Please refer to our preceding publications ^{11 12} for a more detailed description of the experimental setup. 

Generation and Presentation of the Stimuli

Stimuli consisted of full-page texts generated by the same procedure as was used in our previous study on full-page text reading. ¹⁴ Articles were extracted from the Internet Web site of the Swiss newspaper Le Temps (http://www.letemps.ch) and cut into seven-line text segments of approximately 25 words. Arial font (Helvetica) was used. At a viewing distance of 57 cm, the height of the lowercase letter x corresponded to a visual angle of 1.8°. The information content of the stimuli was reduced using one of two pixelization algorithms, square or Gaussian, which differed in the resultant shape of the pixels. These algorithms were applied either off-line, yielding images with “frozen” pixels, or in real-time, yielding “dynamic” pixels that changed with gaze position. 

Square pixelization was performed with a simple block-averaging algorithm, in which matrices of n × n pixels of the original image are fused into single uniform pixels with luminance values corresponding to the mean gray scale levels of the original n × n matrices (Fig. 2a) . 

Gaussian pixelization was performed by applying a two-dimensional (2-D) Gaussian function to each pixel of the stimulus image (Fig. 2b) :  

\[I(x,y)\ {=}\ A({\mu}_{x},{\mu}_{y})\ {\cdot}\ G(x,y).\]

I(x,y) represents the light intensity (gray scale level) at the coordinates (x,y) of the stimulus image. A(μ _x ,μ _y ) is the mean gray scale level of the original n × n pixel matrix with center coordinates (μ _x ,μ _y ). G(x,y) stands for the 2-D Gaussian function calculated as:  

\[G(x,y)\ {=}\ \frac{1}{2{\pi}{\sigma}^{2}}\ e^{\frac{(x\ {-}\ {\mu}_{x})^{2}\ {+}\ (y\ {-}\ {\mu}_{y})^{2}}{2{\sigma}^{2}}},\]

where σ denotes the SD of the particular Gaussian function around its horizontal (μ _x ) and vertical (μ _y ) means. In our case, σ determines the amount of overlap of each pixel onto its neighbors (Gaussian width), whereas μ _x and μ _y correspond to the center coordinates for each pixel (Fig. 3) . 

Off-Line Pixelization.

All text segment images (seven lines of full-page text) used for static presentations were processed off-line, during the preparation phase of the experiment. Subjects could scan these prepixelized images through the 10° × 7° viewing window, under control of their gaze position on the screen. 

Real-Time Pixelization.

In this condition, only the small portion of the entire text segment image displayed in the 10° × 7° viewing window (determined by the subject’s gaze position on the screen) was pixelized in real-time. Gaze position data were used to reposition the viewing window and to display its newly pixelized content on the screen. To achieve adequate image stabilization on the retina, the maximum image-processing time (stimulus pixelization and display) was kept below 10 ms. To fulfill this condition, enormous processing power is needed when large Gaussian widths are used, due to significant amounts of overlap across neighboring pixels. For real-time pixelization, the processing power of our equipment limited us to Gaussian widths up to 0.14 pixels. 

Testing Procedure

The remaining aspects of the experimental procedure were exactly the same as described in our preceding study on full-page text reading. ¹² Briefly, tests were performed monocularly (using the dominant eye) and in central vision. For each run, subjects had to read aloud several text segments of an article, randomly chosen out of a pool of 50 (none of the subjects read an article twice). Test sessions frequently included several runs, but they never lasted longer than 30 minutes, to avoid fatiguing the subjects. 

The programs and algorithms used for image processing and experiment control were developed in commercial software (Visual C++ 6.0 SP5; Microsoft, Redmond, WA) and the latest Platform SDK libraries available at the time of the experiment. Some functions of the EyeLink Windows API library (v. 1.0; SensoMotor Instruments, GmbH) were also used. 

Data Analysis and Statistics

Two variables were measured to assess reading performance: reading scores, expressed in percentage of correctly read words (gender and conjugation mistakes were considered as errors), and reading rates, expressed in the number of correctly read words per minute. Since percentage scales are not adequate for statistical analysis, ³⁰ reading scores were transformed to rationalized arcsine units (rau). Nevertheless, for better clarity, an approximate percentage scale is shown on the right axes of the figures and is also used in the text. 

Results were calculated as the mean of the cumulative performance of each subject ± SEM. Statistically significant differences in reading performance were determined by standard (paired) t-tests with a significance level of 0.05. 

Results

Real-Time Square Pixelization Versus Off-Line Square Pixelization

Five normal volunteers (22, 23, 24, 26, and 28 years of age) were requested to read full-page texts using off-line and real-time square pixelization. Five resolution levels were tested: 28,000, 1,750, 572, 280, and 166 pixels in the viewing window. These resolution levels were identical with those used in our previous study on reading of isolated four-letter words. ¹¹ All subjects started with the easiest (highest) resolution and progressed toward the most difficult (lowest) one. The first four text segments of an article (approximately 100 words) had to be read in each run. Three runs were performed per each pixelization condition. Off-line and real-time pixelization conditions alternated. It is important to note that the first resolution level (28,000 pixels) corresponded to maximum screen resolution (no pixelization had to be performed). Off-line and real-time pixelization conditions were thus identical in this particular case. 

Figure 4compares mean reading performances versus number of pixels in the viewing window for off-line and real-time pixelizations. Individual performances in each experimental condition were established on the basis of 12 text segments and data were fitted with psychometric functions. Down to a target resolution of 572 pixels, average reading scores were close to perfect (above 95% correct) and statistically equivalent for both conditions. At 280 pixels, subjects achieved reading scores of 94.3% with real-time pixelization, but of only 76.4% with off-line pixelization. This difference was statistically significant (P = 0.0017), and persisted at the lowest resolution (166 pixels; 56.1% versus 29.3%; P = 0.013). It is interesting to estimate the critical target resolution for subjects to reach useful reading performances. In our previous study on full-page reading, ¹² we found that adequate (good to excellent) text comprehension correlated closely with high reading scores. This criterion was fulfilled at median scores of 96.8%. In the present case, the fits to the data indicate that this score is reached at 498 pixels in the case of off-line pixelization and at 322 pixels for real-time pixelization (Fig. 4a) . 

Reading rates appeared to be even more sensitive to the number of pixels in the viewing window (Fig. 4b) . At the highest resolutions, subjects reached an average reading rate of 93 words/min. At 572 pixels, mean reading rates had significantly (P < 0.0001) decreased to 80 words/min for real-time and to 64 words/min for off-line pixelization. The difference between both pixelization conditions was also statistically significant (P < 0.0001) and persisted at 280 pixels (34 words/min for real-time pixelization versus 18 words/min for off-line pixelization; P = 0.002). The lowest pixelization condition (166 pixels) was so difficult that reading rates were very low (four to six words/min) in both cases. 

Taken together, these results indicate that equivalent reading performances could be reached at a significantly lower resolution with real-time pixelization. 

Off-Line Gaussian Pixelization Versus Off-Line Square Pixelization

Six normal subjects (26, 29, 29, 33, 34, and 41 years of age) participated in the second experiment. Pixelizations with six different Gaussian widths (σ of 0.036, 0.071, 0.143, 0.286, 0.571, and 1.143 pixels) were tested and compared with square pixelization. The effect of varying the Gaussian width σ for image pixelization is illustrated in Figure 5 . In all conditions, the 10° × 7° viewing window contained 572 pixels (resolution shown to provide enough information for useful full-page text reading ¹² ). Each subject had to read an article of approximately 250 words (i.e., 10 consecutive text segments, per condition). Three subjects started the experiment with Gaussian pixelization at the smallest σ value, progressed toward the larger Gaussian widths, to finish with square pixelization. The remaining three subjects conducted the experiment inversely. 

Mean reading performances versus Gaussian function width (σ) are shown in Figure 6and compared to results obtained with square pixelization. Four Gaussian widths (σ = 0.071, 0.143, 0.286, and 0.571 pixels) resulted in reading scores above 94% correctly read words. These scores were very close to those obtained with square pixelization (Fig. 6a) . Mean reading scores with σ = 0.143 and 0.286 pixels were found to be significantly better than those obtained with square pixelization (P = 0.04 and 0.009, respectively). Reading scores declined markedly below 80% for the two extreme Gaussian widths tested (σ = 0.036 and 1.143 pixels). 

Mean reading rates displayed a similar picture. A maximum reading rate of 70 words/min was achieved at σ = 0.286 pixels. This value is significantly higher (P < 0.001) than the reading rate of 57 words/min achieved with square pixelization. Reading rates with σ = 0.143 and 0.571 pixels were not significantly different from those obtained with square pixelization. For σ = 0.036, 0.071, and 1.143 pixels, reading rates declined markedly (below 40 words/min). 

Taken together, these data reveal that Gaussian pixelization can lead to slightly, but significantly better reading performance than can its square counterpart. This suggests that some degree of image smoothing resulting from overlapping between neighboring pixels can be beneficial for reading. This benefit is, however, only observed for a restricted range of overlapping. 

Real-Time Gaussian Pixelization Versus Real-Time Square Pixelization

Results of the second experiment demonstrated that off-line Gaussian pixelization could lead to significantly better reading performance than off-line square pixelization. A third experiment was thus dedicated to extend this comparison to real-time mode. 

For this evaluation we would have rather used the “optimal” Gaussian width (σ = 0.286 pixels) determined in the second experiment. However, the total processing time needed to simulate this condition turned out to be too important to ensure adequate image stabilization on the retina. Using the second best condition (σ = 0.143 pixels) allowed us to keep processing time below 10 ms. The same six normal volunteers who had participated in the second experiment were requested to read 10 text segments in each of two conditions: (1) real-time Gaussian pixelization at σ = 0.143 pixels and (2) real-time square pixelization. In both conditions, the 10° × 7° viewing window contained 572 pixels. Three subjects started with real-time square pixelization and then switched to real-time Gaussian pixelization. The remaining three subjects performed the experiment inversely. 

The results of this experiment are summarized in Table 1 . No significant difference in performance was recorded between both types of pixelization. However, reading scores and reading rates tended to be slightly higher with square pixelization. Comparing those real-time scores with their off-line counterparts gathered in the second experiment reveals that both real-time conditions yielded better performance. This performance gain was significant for square pixelization (reading scores: P = 0.003; reading rates: P = 0.008), but not for Gaussian pixelization (reading scores: P = 0.12; reading rates: P = 0.25). 

Discussion

The first experiment clearly shows that at low stimulus resolutions (below approximately 1000 pixels in a 10° × 7° viewing area) real-time square pixelization yields better reading performances than its off-line equivalent. The major reason for this performance improvement lies probably in the capability of the visual system to integrate various low-resolution images, enhancing stimulus contrast and resolution ²¹ to improve perception. This effect is also used in standard video: when several low-resolution images are presented in a rapid sequence, the resultant perception is that of a continuous, higher-resolution motion picture. In our experiments, at constant pixel resolution, the readability of pixelized text images depends on the exact position of the pixelization grid relative to the original stimulus image. Therefore, the image can be modified with minor eye movements to optimize viewing conditions. Figure 7illustrates this effect for a series of minor changes in grid position. We observed that subjects quickly adopted this strategy: When resolution decreased, they increased the number of small saccades around the word they were trying to decipher. 

Other effects are also likely to influence reading performance. Previous research on face recognition ^{16 17 18 20 31} revealed that blocked images lead to poorer performance than images filtered using other techniques, mainly because these add artifactual high-frequency components to the target image that may mask essential features for identification. Real-time pixelization does not have the same artifactual bias because pixel movement acts as a low-pass filter that subtracts some of these parasitic frequencies. This could also explain why in the second experiment off-line Gaussian pixelization yielded better reading performance than off-line square pixelization (for a restricted range of Gaussian widths of approximately σ = 0.286 pixels). Additional research, especially at lower resolutions, would be necessary to investigate other factors. It should also be stressed that extreme Gaussian widths noticeably impaired performance. When very small Gaussian widths were used, pixels appeared as isolated small points of light, making it almost impossible to extract a cohesive picture. With large Gaussian widths, overlap was too pronounced, leading to very-low-contrast stimuli. 

Results of experiment 3 might appear surprising in light of the findings of experiment 2: When using real-time processing, the benefits of Gaussian pixelization vanished. In fact, this outcome is not astonishing. Real-time processing had already eliminated the major handicap of square pixelization. The distracting high-frequency noise introduced at pixel borders is low-pass filtered by pixel movement. We believe that the use of the optimal Gaussian width σ = 0.286 pixels (instead of 0.143) would not change this result fundamentally. 

Implications of the Results for Simulations of Artificial Vision

The exact characteristics of the electrophysiological response of the retina to patterned electrical stimulation remain undetermined to this date. However, the use of 2-D Gaussian functions for stimulus pixelization is certainly a more physiologically pertinent approach than the use of square pixels (pixel borders are smoother and it allows for overlapping between neighboring pixels). As soon as the results of electrophysiological experiments on retinal tissue become available, the parameters of such 2-D Gaussian (or more adequate) functions should be adapted. Our experiments also revealed that Gaussian width is an important factor for readability, suggesting that stimulating current strength and electrode spacing might have to be further “tuned” (within safe and comfortable limits) to achieve the most efficient image transmission possible. 

Real-time processing also allows for more realistic simulations of the visual information provided by retinal prostheses. Our results demonstrated that it yields significantly better performance than its off-line counterpart. However, this benefit was relatively moderate, not allowing for a significant reduction (e.g., a factor of two) of the number of stimulation points. Most probably, this advantage will be even less important in visual prostheses with external head-mounted cameras, since head movements are larger and less frequent than eye movements. Recurring head movements could also result in an abnormal vestibulo-ocular reflex. 

The first visual prosthesis prototypes have been recently implanted in humans with encouraging results. ^{5 6 7} Yet, several important challenges still need to be overcome before these devices can provide benefits similar to those of cochlear implants in cases of deafness. The basic notion of patterned vision resulting from the continuous stimulation of several electrodes has not been fully confirmed. An appropriate method of selective stimulation eliciting the adequate psychophysical response has not been developed yet. Another major problem is to achieve efficient electrical stimulation within safe charge density limits. ³² To reduce the total electrical charge injected on the retina, the use of relatively large stimulation electrodes (fundamentally limiting interelectrode spacing) as well as alternate solutions (such as inverted polarity, interleaved stimulation, and/or increasing the total area of the retinal array within feasible limits) may be mandatory. A substantial research effort is therefore still needed to solve these and other open issues before realizing the level of electrode integration suggested by our studies. 

In conclusion, these results demonstrate that the spatial and temporal characteristics of image pixelization play a role in artificial vision simulations. Equivalent performance could be reached with a resolution reduction of approximately 30%, if stimulation parameters were adequate. This effect is not strong enough, however, to change fundamentally the minimum requirements determined in our previous studies on the basis of simplified processing: ^{11 12} Four to five hundred contacts covering a 2 × 3-mm² retinal area are necessary to transmit sufficient visual information for full-page text reading. Reading is particularly important because it is strongly associated with vision-related estimates of quality of life and represents one of the main goals of low vision patients seeking rehabilitation. ^{33 34 35} It is thus important to be aware of such minimal conditions when developing visual prostheses, even if less sophisticated devices might already bring some clinical benefits to patients. 

 Supported by Swiss National Foundation for Scientific Research Grants 3100-61956.00 and 3152-063915.00 and by the ProVisu Foundation.

 Submitted for publication October 4, 2004; revised February 14, May 24, and June 2, 2005; accepted August 1, 2005.

 Disclosure: A. Pérez Fornos, None; J. Sommerhalder, None; B. Rappaz, None; A.B. Safran, None; M. Pelizzone, 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.

 

Experimental setup used for prosthetic vision simulations. Subjects were asked to read full-page texts by using their eye movements to move a stabilized, restricted viewing window on a computer screen.

Figure 2.

 

Pixelization methods: (a) square pixelization (block averaging); (b) Gaussian pixelization.

Figure 3.

 

Gaussian pixelization. A 2-D Gaussian function was applied to each pixel. Block averaging was used to determine the peak of the Gaussian function. σ represents the SD used in the Gaussian function (Gaussian width); μ _x and μ _y are the center coordinates of the stimulus pixel to which the function is applied.

Figure 4.

 

Reading performance versus number of pixels in the 10° × 7° viewing window for five normal subjects. Two stimuli generation procedures are compared in central vision: real-time pixelization and off-line pixelization. (a) Mean reading scores expressed in rau ± SEM (left scale) and in % (right scale). Dashed line: indicates reading scores corresponding to good-to-excellent text comprehension. (b) Mean reading rates expressed in words per minute ± SEM.

Figure 5.

 

Pixelization with various Gaussian widths σ (pixel overlapping). Gaussian pixelizations with: (a) σ = 0.071 pixels (little overlap), (b) σ = 0.286 pixels (medium overlap), and (c) σ = 1.143 pixels (large overlap).

Figure 6.

 

Reading performance versus Gaussian function width (σ) used for stimulus pixelization in six normal subjects. Results are compared with reading performances obtained with square-pixelized stimuli (dashed line, ± SEM). The resolution of the 10° × 7° viewing window in central vision was kept constant at 572 pixels. (a) Mean reading scores expressed in rau ± SEM (left scale) and in % (right scale). (b) Mean reading rates expressed in words per minute ± SEM (left scale).

Table 1.

 

Mean Reading Performances with Real-Time Stimulus Pixelization in Six Normal Subjects

	Gaussian Pixelization			Square Pixelization			P
Mean Reading Scores (rau ± SEM)	115.8 ± 3.6	(99.6%)	117.2 ± 3.4	(99.8%)	0.22 (ns)
Mean Reading Rates (words/min ± SEM)	69 ± 12		74 ± 15		0.35 (ns)

 Gaussian pixelization compared to square pixelization using a 10°×7° viewing area containing 572 pixels.

Figure 7.

 

Illustration of the effect of the initial position of the pixelization grid on the readability of the pixelized word. A single position does not provide enough information to identify the word unambiguously, but by integrating all three of them, the French word “niveau” can be easily recognized.

The authors thank Andrew Whatham, PhD, for insightful contributions and a critical review of the manuscript.