Evaluation of contrast measures in relation to observers perceived contrast

We have carried out a psychophysical experiment to register perceived contrast. 17 observers viewed 15 images, each image was shown for 40 seconds where the observer stated the perceived contrast of the image. The results from the observers indicate that the consensus of contrast among experts decreases as the perceived contrast decreases. Experts also rate the contrast higher then non-experts. A number of contrast algorithms, developed to predict perceived contrast was evaluated against the perceived contrast from the observers. Introduction Contrast can be defined as the difference between the light and dark parts of a photograph. Where less contrast gives a flatter picture, and more a deeper picture. This is only one of the defintions of contrast, others are the difference in visual properties that makes an object distinguishable or just the difference in color from point to point. Because various definitions of contrast is used in different situation, measuring contrast is very difficult. Just measuring the difference between dark and light points of the image does not predicted perceived contrast because perceived contrast is influenced by the surround. Parameters as resolution, viewing distance, enlightment, memory color etc. will all effect how observers perceive contrast. It is clear that contrast is local, but how local does it have to be? A measure of perceived contrast in images is not clearly defined, several measures to predict perceived contrast have been proposed. It is very important in many fields to predict perceived contrast correctly, in image quality assessment and for displays where correct contrast is important. The goal of this research is to evaluate contrast measures, the predicted contrast by the measures are compared against perceived contrast from a psychophysical experiment. Wandell and Zhang [1] found out that S-CIELAB had problems with images with negative contrast (i.e. the luminance value of a point is higher than the local mean of the that point), and they concluded that the ability to predict perceived contrast is very important for image difference models [1]. This is also noted by Wang et al. [2] where the SSIM model incorporate a comparison of contrast to predict image quality. This method has further been incorporated by many researchers [3, 4, 5]. Taylor et al. [6] incorporated a contrast measure in their image fidelity measure. McCann [7] also states the importance of a metric that predict contrast in order to calculate the best image appearance. Peli [8] proposed a local contrast measure (PELI). This important characteristic makes it suitable for the use on natural images. To obtain an efficient measure of contrast, it is necessary to apply the following steps: • The use of a pyramidal structure of band-pass filters (with a width equal to one octave of the bandwidth) centered on different frequencies and distanced one octave from each other. The image is then filtered from the pyramid to obtain a further series of images, each one representing a portion of the image at a prefixed frequency. • The average luminance is calculated at each level frequency. • Every pixel of the image is divided by the average luminance, obtaining a local contrast measure at each level, on a limited bandwidth for every frequency. Tadmor and Tolhurst [9] analysis of contrast (TT) is based on the difference of gaussian model (DOG), modified and adapted to natural images. In the conventional model, the spatial sensitivity in the center of receptive-fields is described by a bi-dimensional Gaussian with amplitude 1.0: Center(x,y) = exp[−(x/rc)(y/rc)] where the radius rc represents the distance beyond which the sensitivity decreases following 1/e with respect to the peak level. The surround component is represented by another Gaussian curve, with a larger radius, rs: Surround(x,y) = 0.85(rc/rs)exp[−(x/rc)(y/rs)]. When the central point of the receptive-field is placed at the (x,y), the output of central component is calculated as: Rc(x,y) = ∑ i∑ jCentre(i − x, j − y)Picture(i, j), while the output of the surround component is: Rs(x,y) = ∑ i∑ jCentre(i − x, j − y)Picture(i, j). The result of the DOG model is obtained as: DOG(x,y) = Rc(x,y)/Rs(x,y). The conventional DOG model assumes that the response of a neuron depends uniquely on the local luminance difference (∆I) between the center and the surround. After the light adaptation process, the gain of the gangliar cells of the retina and the lateral geniculate nucleus (LGN) neurons depends on the average local luminance I. Thus the model response depends on the contrast stimulus. They propose the following three criteria for the measure of contrast: C(x,y) = [Rc(x,y)−Rs(x,y)]/Rc(x,Y ) C(x,y) = [Rc(x,y)−Rs(x,y)]/Rs(x,Y ) C(x,y) = [Rc(x,y)−Rs(x,y)]/[Rc(x,y)+Rs(x,y)] Rizzi et al. [10] proposed a contrast measure (RAMMG) in 2004. This algorithm subsample the image to various levels in the CIELAB colorspace, the undersampling is simple where the image is halfed without pre-filtering. Then calculating local contrast by taking the difference between one pixel and the surrounding 8 pixels, obtaining transition maps of each level. A recombination of the of the averages for each level results in the global measure. This measure was evaluated by changing contrast in different softwares, and comparing the predicted contrast against this. Rizzi et al. propose the RSC algorithm [11], it combines Rizzi et al’s [10] multilevel approach with Tadmor and Tolhurst’s [9] evaluation of a color stimulus. After computing all subsampled images creating a pyramidal image structure starting from the given image, it executes a neighborhood contrast calculation for every pixel in each level using DOG on the lightness CGIV 2008 and MCS’08 Final Program and Proceedings 253 and on the chromatic channels separately. Unlike all other algorithms, in order to consider also isoluminant color contrast configurations, also chromaticity planes of the CIELab space are used, weighted differently than L. This algorithm derives from RAMMG in which the DOG’s substitute the simple neighborood differences. The attempt is to investigate mainly two directions: first checking if the use of DOG’s on the multilevel pyramid have a better performance in considering more extended edges and gradients and second if the use of the chromatic channels in the computation of the perceived contrast lead to more accurate measures. As well as the previous presented measure, only one number of contrast is produced at the end. Calabria and Fairchild [12] carried out an experiment on a set of images changed with different lightness, chroma and sharpness levels. No large differences between experts and nonexperts when it came to rating contrast was found, but there were a larger variability among the non-experts than for experts. It was also identified that observers rated contrast in grayscale images different than for color images, perceived contrast in achromatic images are higher than perceived contrast of very low-chroma images. Experiment Setup 15 different images have been used in this experiment (Figure 1), representing different characteristics. Images 1 and 2 are Figure 1. Images used in the experiment. (a) Image 1 (b) Image 2 (c) Image 3 (d) Image 4 (e) Image 5 (f) Image 6 (g) Image 7 (h) Image 8 (i) Image 9 (j) Image 10 (k) Image 11 (l) Image 12 (m) Image 13 (n) Image 14 (o) Image 15 provided by Ole Jakob Bøe Skattum, image 10 is provided by CIE, images 8 and 9 from ISO 12640-2 standard, images 3, 5, 6 and 7 from Kodak PhotoCD, images 4, 11, 12, 13, 14 and 15 from ECI Visual Print Reference. 17 observers were asked to rate the contrast in the 15 images. 9 of the observers were experts, i.e. had experience in color science, image processing, photography or similar and 8 non-experts non or little experience in these fields. All observers were recruited from Gjøvik University College, both students and employees. Observers rated contrast from 1 to 100, where 1 was the lowest contrast and 100 maximum contrast. The observers were told to rate the contrast as they comprehended contrast, i.e. no definition of contrast was made by the researchers before commencing the experiment. All observers had normal or corrected to normal vision. Each image was shown for 40 seconds with the surrounding screen black, and the observers stated the perceived contrast within this timelimit. The experiment was carried out on a calibrated CRT monitor, LaCIE electron 22 blue II, in a gray room. The observers were seated approximately 80 cm [13] from the monitor, and the lights were dimmed and measured to approximately 17 lux. Results This section contains results from both the psychophysical experiment and from the different algorithms, a comparison between these is also carried out. Perceived contrast Figure 2 shows the perceived contrast stated by the observers with a 95% confidence interval and Table 1 shows mean values and standard deviation for each image. The image rated with the highest mean by the observers is image 15, but it can not be differeniated from many of the other images due to the confidence intervals. The image with the lowest rated contrast is image 13, but this cannot be differenated from a number of other images. Table 1. Perceived contrast results for all observers. Image 15 has the highest mean value, while image 13 has the lowest mean value. Image 15 also has the lowest mean standard deviation, indicating that observer’s concensus about the high contrast in this image. Image Mean value Mean std 1 58,71 19,16 2 57,06 15,42 3 61,76 14,25 4 50,29 23,08 5 70,47 18,69 6 53,94 19,06 7 63,82 16,44 8 57,65 19,13 9 65,00 22,61