Evaluation of perceptual resolution of printed matter (Fogra L-Score evaluation) Thomas Liensberger a, Andreas Kraushaar b a BARBIERI electronic snc, Bressanone, Italy; b Fogra, Munich, Germany ABSTRACT A new method has been developed to address the objective evaluation of perceived resolution of printed matter. To achieve this, a psychophysical experiment has been designed and conducted that ranked typical test prints representing the manifold of printing processes and substrates. A scanner based method has been developed that computes a score value between 0 and 100 termed Fogra L-Score. It is based on the idea to identify a predefined signal in an image (print). The predefined signal is a perfect representation of the perceived resolution domain by means of a test target (RIT ConRes-target) that covers systematic variations of the two governing parameters of perceived resolution namely contrast and spatial resolution. The printed images to be evaluated have been scanned and pre-processed. The level of closeness between the reference (digital representation) and the printed matter (scanned print) have been determined by a 2-dimensional normalized cross-correlation (on the CIEL*-channel). The resulting correlation coefficients have been compared against findings of the performed psychophysical experiment. Finally a framework will be presented that also allows for spatial filtering to address different intended viewing distances as well as chromatic test charts. Keywords: Perceptual resolution, objective method, computer aided, scanner qualification, 2D cross correlation, L- Score, Matlab 1. INTRODUCTION Image quality has been researched for more than hundred years. It is defined as an impression of the overall merit or excellence of an image, as perceived by an observer neither associated with the act of photography, nor closely involved with the subject matter depicted. In the realm of conventional or digital printing the image quality can be reduced to printed matter and is known as print image quality. Although there is a desire for single values, it must be said that so far no single values have been found that correlate with the overall print image quality. Hence most of the published image quality work concentrates on individual aspects for a defined image attribute and a defined use case, material combination or market segment. A set of different image quality attributes and their correlation have just been evaluated in [1]. In addition, the Image-Quality-Circle of Peter Engeldrum [2] makes it possible to distinguish between technology variables, physical image attributes and subjective perceptions like graininess, homogeneity and resolution [3]. 1.1 Perceived image quality The evaluation of perceived image quality in prints is an active field of research. Definitions of print quality attribute measurements that correlate with visual perception by technology independent means, even across many printing technologies, is under current scrutiny. Image quality is influenced by a number of different quality attributes that are, for convenience only, categorized into color and surface finish, homogeneity, perceptual resolution and artifacts [4]. ISO 29112 [5] says that printer resolution, a quantification of a digital printing system s ability to depict fine spatial details, is a perceptually complex entity with no single, simple, objective measure. It further defines 5 print quality characteristics: native addressability, effective addressability, edge blurriness, edge raggedness, and the printing system modulation transfer function (MTF) that somehow contribute to the perceived resolution. The difference between perceived resolution and physical addressability (often also termed as resolution) is shown in Figure 1.
Figure 1. Difference between physical and perceived resolution. The perceived resolution in both pictures seems to be nearly the same. A big difference is seen in the physical resolution (highlighted boxes). Today a variety of approaches to assess resolution is available, but all of them are based on visual estimations and there exists no method to do this in an objective manner [6]. This paper describes a novel method to evaluate the perceptual resolution by a computer-aided method and to benchmark an image by an index (in a range between 0 and 100) called Fogra L-Score. 1.2 Judging perceived resolution Previous research at Fogra indicated, that a test target comprising a reasonable amount of patches, each of which contains spatial patterns (e.g. concentric circles) of varying line width and lightness contrast, provides a very high correlation with the perceived resolution [7]. Because of fulfilling all this criteria the RIT Contrast-Resolution Target was selected to be used for the present evaluation method (Figure 2). Figure 2. Microphotograph of the RIT Contrast-Resolution Test Target V. 1.8 computed for an addressability of 600 dpi. The 100 circles are defined as vectors (line work). In the vertical direction, there are columns with logarithmically spaced line frequencies ranging from 6.25 to 0.625 cycles per millimeter. In the horizontal direction there are rows with logarithmically spaced contrasts ranging from 100% to 1%. Column A contains all circles with a contrast of 100%.
2. PSYCHOPHYSICAL EXPERIMENTS Getting a calibration dataset for the algorithm was the first step for setting up the experiments. 20 of more than 200 printouts produced on different printing systems covering offset lithography, electrophotography and inkjet have been selected. These printouts had to cover all visual quality levels from bad to perfect on different materials like paper, canvas, textile and roll-up. In addition, a set of five printouts have been selected as the verification dataset to check the robustness of the algorithm by the same experiment and a final rank order experiment. The performed psychophysical experiments have been conducted with both of these datasets under stable and reproducible conditions (D50 daylight with an intensity of 2000lux and a viewing distance of 50cm). As observer have been selected persons with previous experience in image analysis because the estimation of resolution is performed mostly by experts in this area. The use of glasses and contact lenses was allowed. The question asked to the observer was: Mark and count all the circles you can see without distracting artifacts such as aliasing patterns. The method of single presentation has been selected, which means that printouts have been shown one by one and could not be compared next to each other (no reference method). Every one of the observer had to evaluate 20 different printouts. Four of them, which were randomly selected, have been shown two times to check for consistency. The inter- and intra-observer variability was found to be very consistent. The correlation of the so derived index (counting all patches with non-distracted content - from 0 worst to 100 perfect ), called visual Counting Index (visual CI), correlated very well with the results of a rank order experiment based on more than 50 print samples asking to sort all prints (focusing on pictorial content, see Figure 3) by their highest perceptional resolution, i.e. details sharpness. Figure 3. Screenshot of page 2 of the Fogra image quality test form. The contrast-resolution test form of Franz Sigg can be found in the middle of the page. 3. METHOD DESCRIPTION The underlying concept of proposed measurement method is based on an idea from audio restoration. For de-clicking algorithms often a representative click (e.g. the result of a physical scratch) is captured followed by a comparison of the entire record with the reference click through a 1 dimensional cross correlation. With respect to the measurement of perceived resolution the reference is the well-defined representation of systematic variations in visual contrast and spatial resolution. This test chart will be printed and is then subject for evaluation namely a two dimensional cross correlation. Additional steps are included to allow for practical reasons. They are shown in Figure 4.
Figure 4. Printing and scanning (1) the digital Testform are the first two steps to perform. All further steps are computer aided and therefore part of the MATLAB evaluation method. Preprocessing (2) the image and the following spatial filtering (3) prepare the scanned image to be compared (2D cross correlation) with the spatial filtered reference test form (4). Tresholding (5) the correlation coefficients and summarizing the resolved patches leads to the counting index (CI). This is a value between 0 and 100 which is the base to calculate the final L-Score (6). 3.1 Scan In order to evaluate a printed image it is vital to capture it without adding unnecessary errors. Therefore a scanner that produces high quality results needs to be used. However it is normally not a measuring device hence colorimetrical calibration and profiling is required (e.g. ICC scanner profile). This ensures to capture the luminance rather than relying on the default red, green and blue channel signals. Color correction or filters have to be disabled. In addition the optical path should also be considered by applying the reverse MTF of a printer to the scanned image. In this work a high quality scanner was used (Linoscan XL2400). The concrete scanner qualification is still subject for final standardization and will be addressed in ISO/TS 18621-31 (a rough scanner qualification method can be found in document [4]). The scanned printouts are preferably saved as CIELAB-TIFF files using 1200dpi. Moreover bidirectional illumination is needed to compensate material roughness and shadows that could be caused by a unidirectional illumination. It would lead to issues calculating the final detail sharpness value. To avoid interferences (e.g. Moiré effects) on scanned images the printouts have to be scanned parallel to the scanner axes. 3.2 Preprocessing Passing different checks (size, rotation and distortion) was a condition of scanned printouts to be admitted for further processing. Resizing or distorting after capturing the images is not permitted because of causing artifacts and interpolated pixel. Only right sized and paraxial oriented images are admitted to the next evaluation steps. Image registering is performed by a 2D cross correlation which allows finding the exact overlapping position of the scanned image and the reference. This is essential for the performed cross correlation (step 3.4) between each of the 100 patches with the reference patches. Only perfectly overlapping patches permit correct correlation results. Minor rotations of the image have no influence to the result because of the concentric patch content.
3.3 Spatial Filtering Spatial filtering is applied to reproduce the effects of the visual system viewing images from a define distance (the visual angle). This work concentrates on viewing under reading distance, which is approximately 40 to 50 cm. Scanning images means getting the picture content in a fixed resolution including all spatial details. This is more than a human eye can see in normal reading distance and therefore a solution had to be found to simulate blurring in the scanned images. In other words, the image has to be saved like an impression captured by a human eye in normal reading distance of 50 cm. A well-dimensioned Gaussian filter solves this problem by flattening pixels and eliminating all structures and details which are not visible from 50 cm. This filter was applied to the scanned printouts and the reference file which is used to be compared with the scanned images in step 3.4. Figure 5. Printout before and after applying the Gaussian-blurring filter with a radius of 10px (viewing distance simulation for a human eye). 3.4 Normalized two-dimensional cross correlation The core part of this step is the calculation of the normalized correlation coefficient between every patch in the reference image and the corresponding patch in the scanned image. This leads to a set of 100 correlation coefficients, which will undergo a threshold in the next step. A correlation is a mathematical operation that permits finding the similarity of two signals. A perfect accordance of a signal with its reference signal results a correlation value of 1 (when a normalized correlation function is used). The correlation C will be computed for all 100 patches with = {1,,100}. The reference patch (template) is denoted as. The mean of the template is and the scanned print out (image) is. ( ) [ ( ) ] [ ( ) ] [ ( ) ] [ ( ) ] (1) is the mean of ( ) in the region under the template. 3.5 Thresholding These 100 correlation coefficients were subject to be categorized in order to reflect the visual findings. Different thresholds were tested. The best results were found by simply marking patches with C 0.5 as not resolvable while patches with a correlation coefficient of C > 0.5 are marked as resolvable. The first patch with C 0.5 in a column is crucial when analyzing the correlation coefficient of each patch (starting on top of every column). This patch indicates the first occurrence of a non-resolvable patch, which means that all following patches must also be marked as not resolvable to exclude correlation errors produced by the disappearance of resolution. Summing all resolvable patches results in the so called counting index CI with CI = {0,,100}. CI = 0 means that no patch can be resolved and CI = 100 means that all patches can be resolved.
3.6 L-Score computation Finally the L-Score had to be calculated (by means of the counting index) and fine-tuned by comparing it with the findings of the visual experiments (L-Score of visual data - ground truth). The details of the visual experiments are outlined in section 2. The number of the visually resolved patches (termed visual counting index) varied from 23 to 63. In the light of an easy to understand metric, ranging from 0 (worst detail sharpness) to 100 (perfect detail sharpness), a non-linear transformation was applied. In this work a cumulated Gaussian function with two constants (µ=45, σ=12) performed best. This transformation supported the observation that the difference between a CI value of 80 or 100 (also 0 and 20) is not easily resolvable. The L-Score formula with C1 and C2 as variables is the following: ( ) (2) The two constants C1 and C2 are needed to tune the final formula to the visually determined L-Scores (Fogra L-Score of visual CI with C1 = 0 and C2 = 0). Through adjusting C1 the correlation between the visual and computer calculated L- Scores can be minimized. The influence of this constant is a bending over the whole range of values. In this way, it is possible to adapt the trend line of the visual scores to those of the computer calculation. Based on tests an optimal C1 value of 3.89 could be calculated, which increases the correlation to 0.9729. The second constant C2 is an offset value, which permits to minimize the cumulated absolute offset between visual and computer calculated L-Scores. This constant deals only with the adjustment of values on one axis and doesn t influence the correlation. The minimal mean deviation of 4.38 per target can be reached with a C2 value of 2.30. The final formula to calculate the L-Score is the following: ( ) (3) 4. RESULTS The goal of this paper was to find an objective method to estimate the perceived resolution of printouts, including different influencing factors from the environment and the human eye. The differences between the visual L-Scores and the Matlab L-Scores can be seen in Figure 6 and Figure 7. The L- Scores in Figure 7 result by using a wrong set of values for the constants C1 and C2 for the computer calculated Scores. By optimizing the correlation and minimizing the cumulated offset of entire data a correlation of 0.973 is reached. This value can be considered reasonably high.
Fogra L-Score Fogra L-Score 100 90 80 70 60 50 40 30 20 10 0 visual evaluation MATLAB evaluation Calibration printouts Figure 6. L-Score evaluation and matching with visual L-Scores using non optimal constants (C1 = 10, C2 = 0) for computer calculated L-Score. Visual L-Score is always calculated with C1 = 0 and C2 = 0. 100 90 80 70 60 50 40 30 20 10 0 visual evaluation MATLAB evaluation Calibration printouts Figure 7. Visual and Matlab L-Score evaluation with right constants in the L-Score Formula for computer calculated values (C1 = 3.89, C2 = 2.30). The Fogra L-score correlates very well with the visual evaluation of the user experiments. To evaluate the robustness and accuracy of the algorithm a set of visual experiments and Matlab evaluations were performed. For that five verification prints were used, which have not been used to calibrate the algorithm. The results of this final test are visualized in Figure 8. In addition, a rank order experiment using the same five verification prints was performed by the same participants. The goal was to check the accordance between the visually estimated L-Score and the detail sharpness of the pictorial content ordering the test forms by focusing on perceptual resolution of the pictorial content. The results of all participants were the same: an ordered dataset corresponding to the ordering by L-Score values.
Fogra L-Score 100 90 80 70 60 50 40 30 20 10 0 178 215 206 219 221 Verification printouts visual evaluation MATLAB evaluation Figure 8. Results of the verification test. A correlation of 0.985 and a mean deviation of 4.45 between visual and Matlab score can be seen as recently high. Fogra L-Score 90 80 60 40 20 Perceived resolution Perfect Very good Good Satisfactory Adequate < 20 Poor Figure 9. Different L-Score levels of printed matter. 5. SUMMARY AND DISCUSSION Digital printing and the evaluation of perceived image quality of prints is an active field of research. Definitions of measurements of print quality attributes that correlate with visual perception by technology-independent means, even across many printing technologies, is under current scrutiny. In this paper a novel method has been presented that correlates well with the perceived resolution (also known as detail sharpness or resolving power). It is based on the idea to identify a predefined signal in an image (print). The predefined signal is a perfect representation of the perceived resolution domain of a test target that covers systematic variations of the two governing parameters of perceived resolution namely contrast and spatial resolution. The printed image to be evaluated has been scanned and pre-processed. The level of closeness between the reference (digital representation) and the printed matter (scanned print) has been determined by a 2-dimensional normalized cross-correlation (on the CIEL*- channel). The resulting correlation coefficients have been compared against findings of a psychophysical experiment to be conducted beforehand. Finally a cumulated Gaussian distribution has been used to optimally fit the psychophysical dataset resulting in an easy to communicable score (L-Score) value ranging from zero to 100 (perfect detail sharpness). The result of this work was a MATLAB algorithm to estimate the perceptual resolution of digital prints in an objective manner. It has been reached a correlation of 0.985 and a mean deviation of 4.45 between MATLAB calculated and visually estimated L-Scores using a set of randomly selected test charts.
ACKNOWLEDGEMENTS The project Evaluation of perceptual resolution of printed matter (Fogra L-Score evaluation) is a cooperation project between Fogra (Germany), BARBIERI electronic (Italy) and the University of Technology Vienna (Austria). REFERENCES [1] Pedersen, M., Bonnier, N., Hardenberg, J. Y., Albregtsen, F., "Attributes of image quality for color prints," Journal of Electronic Imaging 19 (2010). [2] Engeldrum, P. G., [A Theory of Image Quality: The Image Quality Circle], IS&T - The Society for Imaging Science and Technology, Winchester, Massachusetts, USA (2004). [3] ISO/CD, "ISO/CD 15311 - Graphic technology - Requirements for printed matter utlilizing digital printing technologies for the commercial technologies and industrial production" (2011). [4] Liensberger, T., "Objektive Bewertungsmethoden zur Ermittlung der Detailschärfe von Drucken", Vienna (2013). [5] ISO/IEC, "ISO/IEC 29112 - Information technology - Office equipment - Test charts and methods for measuring monochrome printer resolution" (2009). [6] Sigg, F., "Testing for Resolution and Contrast", R.I. School of Print Media (2006). [7] Komischke, C., "Bewertungskriterien der Auflösung von Digitaldruckmaschinen und betriebswirtschaftliche Implikationen eines Digitaldruckstandards", Wismar (2011).