Analysis of the subjective quality estimation of photo prints with reduced jpeg qulity.

Bota, Josip ; Milcic, Diana ; Donevski, Davor 等

1. INTRODUCTION

The accelerated development of digital electronic devices influences our daily routine. Today, a personal computer and a digital camera make a part of the inventory of almost every household. Digital photography became a standard, replacing the conventional photography. Due to its simplicity, using it does not require a great deal of technical knowledge. Differences in camera quality (optics, sensors, file format, etc.) and computer processing determine the final image quality. The growing demands for transferring greater amounts of data push hardware solutions to their limits. That is the point where software solutions come into place, attempting to compress the date so the same information could be transferred in less time. Transfer of large image files is the bottleneck of distribution systems, and therefore every reduction of file size represents a success for saving both time and storage space.

Image compression is a vital instance in multimedia communications (Singh et al., 2007). Deviations in color reproduction appear in every color reproduction process. The two important instances about compression are its platform dependence and performance. Today's compressions are mostly platform independent because they conform to international standards. Human evaluation of color and quality is highly individual, and calls for the determination of the right quality level satisfying most users' needs.

2. THEORETICAL

JPEG is a standard procedure for image compression. The Joint Photographic Experts Group is the standard author. JPEG was created for compression of both monochrome and color images, and is used for photographs, natural painting and similar applications. It is not suited for text, simple drawings and technical drawings. JPEG compression uses the flaws of the human eye, i.e. the fact that human perception is more sensitive to small differences in lightness than the differences in color (Kayand & Levine, 1994). JPEG compression uses the DCT (Discrete Cosine Transform), and divides its coefficients by their corresponding quantum. Figure 1 shows the block diagram of JPEG compression. It should be noted that different image processing applications use different quantization tables. Therefore, the results achieved at same compression levels differ, and we cannot express any recommendations as certain compression level values. Approximate quality factors are better indicator and should be considered for this purpose.

[FIGURE 1 OMITTED]

3. EXPERIMENTAL

3.1 Hypothesis

The subjective human estimation of the compressed images quality is inferior to the objective estimation. This brings the importance of optimizing the reproduction process. It is known that digital images using greater JPEG compression suffer greater amounts of loss in details. This particularly relates to small differences in color on more or less uniformly colored parts of the image. Based on the experience as a photographer and a graphic designer, I expect that one part of the examinees will not notice any differences in quality of the images.

It is assumed that the reproduction quality does not differ significantly among different photo studios due to the use of standard substrates and printing machines of similar technical characteristics. The assumption is based on the equable market price and opens the possibility of determining the needs and the cost effectiveness on our market and most likely on the global market as well.

3.2 Materials and methods

The determination of subjective estimations requires a template containing familiar objects with as wider variety of chromaticity and luminance points.

One of the color consistency control tools is Gretag-Macbeth ColorChecker Chart. It allows the control of color reproduction by measuring the values of the 24 reference patches. The colors of the patches are chosen to allow the tracking of the most common colors which appear in photography (flowers, light and dark human skin, etc.).

The examination template and the ColorChecker Chart are combined in the same file. The file was saved using the image processing software in compression levels 7 to 12. The reference level is 12, and the comparison with levels 10, 8, 6, 4, 2, and 0 was made. For the purpose of collecting the data, every level was assigned a letter as an indicator (Table 2). Those photographs were printed in photo studios, two samples for each compression level. The overall template format was 1772 x 2488 pixels, 300 dpi or 15 x 21 cm in size.

On the delivery of files, the photo studio personnel were instructed to print the photographs on a glossy photographic paper in a given format without any corrections. Such corrections, commonly made in photo studios, would result in uneven conditions of producing the photographs, and incomparable results for the analysis.

The photographs were later detached from the ColorChecher Chart so it wouldn't affect the evaluation. 50 examinees of different age and sex analyzed the templates in different conditions, and marked those on which they noticed difference in quality compared to the reference. The chart patches were measured using a spectrophotometer with standard illuminant D50, and measured values compared to the referent values. The differences are expressed as [DELTA][E.sub.00].

3.3 Results

Table 1 shows the [DELTA][E.sub.00] difference between the reference print (A) and for the rest for each patch. In addition to [DELTA][E.sub.00] for each patch, mean values of [DELTA][E.sub.00] for each print are presented in Table 2. Table 3 summarizes the analysis results. The number of examinees which notices reduction in compressed images quality compared to the reference is expressed as percentage.

3.4 Discussion

The analysis of the measurement results shows that images with greater compression levels suffer greater deviations. There is no specific patch that suffers greater deviation than the other patches. Mean value of [DELTA][E.sub.00] and relatively small data dispersion indicate that greater compression levels result in greater values of [DELTA][E.sub.00]. These results are especially noticeable in prints "E" and "G".

50 examinees evaluated 6 prints comparing them to the reference in different conditions. The examinees are not graphic technology experts, nor did they evaluate the prints under standard lighting conditions. The examination resulted with the following answers: 6% for print "C", 26% for print "G", 54% for print "E" i.e. that was the percentage of the examinees who noticed the difference in the print quality compared to the reference, while 46% did not notice any difference. Non expert examinees and non standard viewing conditions were used to inspect what results can be expected with the majority of population and most common viewing conditions. Comparing the file sizes (Table 2) reveals that they differ significantly. Comparing the files "A" and "B" reveals that file "B" is approximately 1/3 the size of file "A". Comparing "B" to "F" reveals that file "F" is approximately 1/2 the size of file "B". Other files, "D", "C", "G", "E", are more similar in size.

4. CONCLUSION

The investigation determined that using JPEG compression which is adapted to the flaws of the human eye, and using the premium quality digital printing, 46% of the examinees did not notice any difference between the samples. That indicates that the examinees either were not able to recognize the difference, or that the difference was not in significant factors which would draw their attention. The 54% noticed the difference, but only 6% noticed the difference at the compression level 4. Considering that the examinees compared the samples directly, there is still an open question if the results would differ if they had compared the samples by memorizing them. The results can be used as guidelines for users to use the greater compression levels safely, which most of them usually avoid. The results show that the viewers estimated that the best ratio between quality and compression level is achieved at compression level 8, which corresponds to approximate quality factors of 88,28 for luminance, and 90,19 for chrominance. At that compression level, the sample image file was 1/4 the size of the original, and was not on the borderline where viewer starts noticing reduction in quality.

This paper calls for further investigations of factors influencing the subjective estimation of quality. The viewing conditions, age and sex are all factors which influence the subjective estimation to higher or lower extent.

5. REFERENCES

Aldaba A., M.; Linhares M.M., J.; Pinto D., P.; Nascimento M.C., S.; (2006). Visual sensitivity to color errors in images of natural scenes, Visual Neuroscience, United Kingdom, printed in the USA

Alkholidi, A.; Alfalou, A.; Hamam, H.; (2007) A new approach for optical colored image compression using the JPEG standards, Signal Processing, Elsevier North-Holland, Inc. Amsterdam, The Netherlands

G.K, W.; (1991). The JPEG Still Picture Compression Standard, Communications of the ACM, ACM New York, NY, USA

K.M. Au, N.F. Law, W.C. Siu, (2007). Unified feature analysis in JPEG and JPEG 2000-compressed domains, Pattern Recognition, Elsevier Science Inc. New York, NY, USA

Kayand C., D.; Levine R., J.; (1994). Graphics File Formats, Windcrest, McGraw-Hill

Singh, S.; Kumar, V.; H.K. Verma, (2007). Reduction of blocking artifacts in JPEG compressed images, Digital Signal Processing, Academic Press, Inc. Orlando, FL, USA

Table 1. [DELTA]E of the prints compared to the reference

 [DELTA] [DELTA] [DELTA]
 [E.sub.00] [E.sub.00] [E.sub.00]
Patch B C D

 A1 0,41 0,45 0,32
 A2 0,49 2,79 0,67
 A3 0,55 1,84 0,17
 A4 0,29 0,27 0,75
 A5 0,36 1,16 0,30
 A6 0,83 0,67 0,92
 B1 0,31 0,40 0,40
 B2 0,63 0,30 0,72
 B3 0,17 0,34 0,43
 B4 0,20 0,32 0,21
 B5 0,30 0,38 0,21
 B6 0,27 0,19 0,17
 C1 0,49 0,51 0,70
 C2 0,13 0,58 0,34
 C3 0,12 0,95 0,35
 C4 0,33 0,68 0,56
 C5 0,44 0,93 0,42
 C6 0,30 0,23 0,32
 D1 0,84 1,95 1,04
 D2 0,45 1,18 0,61
 D3 0,44 0,73 0,66
 D4 0,53 0,55 0,49
 D5 0,66 0,64 0,88
 D6 0,94 2,03 0,95

 [DELTA] [DELTA] [DELTA]
 [E.sub.00] [E.sub.00] [E.sub.00]
Patch E F G

 A1 10,90 0,31 0,54
 A2 11,27 0,53 2,02
 A3 11,03 0,41 2,06
 A4 10,48 0,42 0,85
 A5 7,54 0,60 1,29
 A6 5,15 1,00 0,87
 B1 6,61 0,13 0,48
 B2 10,46 0,63 1,23
 B3 7,89 0,25 0,71
 B4 11,09 0,24 0,46
 B5 10,53 0,15 0,64
 B6 10,12 0,52 0,28
 C1 10,43 0,37 1,59
 C2 8,87 0,23 0,67
 C3 10,62 0,47 1,36
 C4 5,57 0,39 1,05
 C5 10,49 0,64 0,46
 C6 10,26 0,57 0,56
 D1 7,92 0,60 1,67
 D2 10,31 0,50 1,15
 D3 10,92 0,43 0,73
 D4 9,35 0,79 1,48
 D5 11,13 0,63 1,97
 D6 10,61 0,95 2,38

Table 2. Mean [DELTA][E.sub.00]

 B C D E F G

Mean [DELTA]
 [E.sub.00] 0,43 0,83 0,52 9,56 0,49 1,10

Table 3. Subjective estimation

 B C D E F G

Noticed 0% 6% 0% 54% 0% 26%
difference