CC BY
0
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
JPEG and BPG visually lossless image compression via KonJND-1k database
Boban P. Bondzulic3, Nenad M. Stojanovicb,
Vladimir V. Lukinc, Sergii S. Kryvenkod
а University of Defence in Belgrade, Military Academy, Department of Telecommunications and Informatics, Belgrade, Republic of Serbia, e-mail: bondzulici@yahoo.com,
ORCID iD: https://orcid.org/0000-0002-8850-9842
b University of Defence in Belgrade, Military Academy, Department of Telecommunications and Informatics, Belgrade, Republic of Serbia, e-mail: nivzvk@hotmail.com, corresponding author,
ORCID iD: https://orcid.org/0000-0001-9328-5348
c National Aerospace University, Department of Information-Communication Technologies, Kharkiv, Ukraine, e-mail: lukin@ai.kharkov.com,
ORCID iD: https://orcid.org/0000-0002-1443-9685
d National Aerospace University, Department of Information-Communication Technologies, Kharkiv, Ukraine, e-mail: krivenkos@ieee.org,
ORCID iD: https://orcid.org/0000-0001-6027-5442
doi https://doi.Org/10.5937/vojtehg72-50300
FIELD: telecommunications ARTICLE TYPE: original scientific paper
Abstract:
Introduction/purpose: This paper presents the results of the research on visually lossless image compression which is of particular interest because it achieves a high degree of compression, while the visual quality of the image is not impaired, i.e., end users are very satisfied with the image quality. The analysis was carried out using the publicly available large-scale picture-wise KonJND-1k database which contains the results of subjective tests on JPEG and BPG compressed images.
Methods: Thanks to the availability of images from the KonJND-1k database, the dependence of objective assessments of image quality on parameters that control the degree of compression of source signals (quality factor for JPEG and quantization parameter for BPG) is analyzed. The results of the visually lossless subjective tests are used for a deep analysis of the boundary and typical values of the parameters that control these two types of compression, as well as for the analysis of the corresponding values of the objective quality scores. Furthermore, reliable features for predicting the boundary between visually lossless and visually lossy compression have been identified. For that purpose, the degree of agreement between the predictions and the ground truth values of the peak
signal-to-noise ratio (PSNR) and image representation in bits per pixel (bpp) is used. The visually lossless compression ratio is used to compare JPEG and BPG techniques.
Results: It is shown that the boundary between visually lossless and visually lossy image compression is found in a wide range of PSNR values (about 20 dB for JPEG and 15 dB for BPG). The corresponding JPEG image compression quality factor values at this threshold also range widely from 31 to 79, with concentration between 40 and 45. For the BPG encoder, the values of the quantization parameter are grouped around 30, and the boundary values are 25 and 34. Furthermore, it is shown that this boundary can be reliably determined based on simple features derived from the original uncompressed image. Gradient-based features known as spatial frequency and spatial information proved to be the best predictors. The degree of agreement between the predictions obtained from these features with the ground truth values of PSNR and bpp in both types of compression is greater than 85%. A comparative analysis has showed that, using BPG compression, it is possible, on the average, to achieve a twice larger compression ratio of visually lossless compression than for JPEG (80 versus 40).
Conclusion: Although a high degree of agreement is achieved between the predictions and the ground truth values of PSNR and bpp of the boundary between visually lossless and visually lossy compression, there is a need for the development of new prediction approaches, especially with the BPG technique, which through the compression ratio proved to be superior to the JPEG technique. The existing databases used for the analysis of visually lossless compression contain color images from the visible part of the electromagnetic spectrum. Considering the increasing use of images from the infrared part of the spectrum, there is a need to conduct similar tests in this spectral range.
Key words: BPG compression, JPEG compression, just noticeable difference (JND), peak signal-to-noise ratio (PSNR), visually lossless image compression.
Introduction
With the increasing use of images and videos in everyday life, compression of visual signals is gaining more and more importance. The compression process adapts the image/video to the bandwidth of the transmission system while the requirements for memory resources are reduced. It is well known that compression techniques can be divided into lossless compression and lossy compression (Bull & Zhang, 2021). In real applications, the degree of compression of lossless techniques is often not sufficient, while lossy compression techniques can achieve a significantly
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
higher degree of compression, but at the expense of impairing the visual quality of the signal. As a compromise solution, visually lossless techniques have been increasingly used in recent years. Some researchers classify these techniques as the third type of compression, while some of them divide lossy compression techniques into visually lossless and visually lossy (Krivenko et al, 2018).
The just noticeable difference (JND) concept is very important within the visually lossless techniques, and it refers to the smallest degree of degradation that the observer is able to notice if the image is compressed with some compression process (Kovalenko & Lukin, 2023). The JND threshold can be defined for individual pixels (pixel-wise), at the region level (patch-wise), or at the level of the entire image (picture-wise) (Shen et al, 2021).
Thanks to the publicly available JND image databases with subjective test results, such as MCL-JCI (Jin et al, 2016), JND-Pano (Liu et al, 2018), JND-VVC (Shen et al, 2020), VLT (Mikhailiuk et al, 2021), KonJND-1k (Lin et al, 2022), researchers are enabled to analyze the JND concept in different types of compression. The JPEG compression type has been included in MCL-JCI and JND-Pano databases, JND-VVC database uses Versatile Video Coding (VVC) compression, VLT database contains JPEG and WebP compressed images, and the large-scale crowdsourced JND database KonJND-1k contains test results on JPEG and Better Portable Graphics (BPG) compressed images. The JPEG standard was developed more than 30 years ago and is still widely used, while the BPG coder is one of the promising lossy compression coders (Bellard, 2018; Kovalenko et al, 2022). The degree of JPEG compression is controlled using a quality factor (QF) that ranges from 1 to 100, where lower QF values provide a higher degree of compression. BPG compression uses a quantization parameter (QP) that ranges from 0 to 51, where lower values correspond to better visual quality (Lin et al, 2022). As BPG relies on the H.265/HEVC video compression standard, this dynamic range corresponds to the QP values used to adjust the quality of the H.265/HEVC compressed video signal. Recently introduced JND databases include subjective tests on a larger number of codecs (Testolina et al, 2023), and on a larger number of degradation types (Liu et al, 2023).
The boundary between visually lossless and visually lossy image compression (first JND point or JND #1) can be determined through QF/Qp prediction, image representation prediction in bits per pixel (bpp) or through objective quality prediction using some of the measures such as PSNR (Saha & Vemuri, 2000), PSNR-HVS-M (Zemliachenko et al, 2016), SSIM (Cai et al, 2019), FSIM (Li et al, 2021), MDSI (Li et al, 2022),
VIF (Fiorucci et al, 2012) and similar measures. The simple approach presented in (Bondzulic et al, 2021) to predict the PSNR of the first JND point of JPEG compressed images uses only one feature of the original uncompressed image (the mean gradient magnitude). The effectiveness and acceptable prediction error of the PSNR value of the first JND point of this approach inspired the authors in (Pavlovic et al, 2023) to analyze the prediction based on other features, where it was shown that good predictors are spatial information, spatial frequency, contrast and degree of compression. Deep learning approaches (Fan et al, 2019; Lin et al, 2020; Liu et al, 2020) for the prediction of the position of the first JND point of JPEG compressed images provide a smaller PSNR prediction error and can be used for direct QF/QP prediction.
The proposed approaches, where the prediction of the position of the first JND point is carried out through bpp or objective measures, require that the desired bpp or objective quality be reached through an iterative way by changing Qf/QP (Zemliachenko et al, 2016). In the literature, one can find approaches such as the two-step approach (Li et al, 2024) or the approach with linear interpolation of the rate distortion curve (Poth et al, 2020; Bondzulic et al, 2024) by which the process of achieving the desired quality is carried out in a few iterations. However, the error introduced by such approaches is superimposed on the prediction error of the position of the first JND point.
In this paper, the results of subjective tests of the KonJND-1k database are analyzed, as it is the largest JND database, which covers JPEG and BPG compressed images, and instead of tests in a laboratory environment, the authors used a crowdsourcing framework. The positions of the JND points in both types of compression are analyzed through the values of PSNR (Wang & Bovik, 2009), PSNR-HVS-M (Ponomarenko et al, 2007) and bpp. Dependences of objective quality measures on parameters that control the degree of compression are shown. The possibility of predicting objective quality scores of visually lossless compressions based on simple features derived from original uncompressed images is analyzed. Finally, the gain (compression ratio) that can be achieved with both types of compression at the JND threshold is given.
KonJND-1k database content analysis
KonJND-1k is a publicly available picture-wise JND image database (Lin et al, 2022). The database consists of 1008 source images and their JPEG and BPG compressed versions with dimensions of 640x480 pixels,
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
together with the results of JND subjective tests. One half of the source images is compressed with JPEG, while the other half is compressed using the BPG encoder. The degree of compression of JPEG images is controlled by the quality factor (QF) which ranges from 1 to 100, so that for one original image there are 100 JPEG compressed images. The degree of compression of BPG images is controlled by the quantization parameter (QP) which is changed from 1 to 51, so that there are 51 BPG compressed images per one original image.
A total of 503 observers participated in the subjective tests, and an average of 42 opinions per original image were collected. Unlike the MCL-JCI (Jin et al, 2016) where the subjective results were collected using binary search comparison, in the KonJND-1k database the results were collected using slider adjustment and a flicker test to determine the position of the first JND point. In the flicker test, the original image and its compressed version alternate at a frequency of 8 Hz, whereby subjects drag the slider to the position corresponding to the smallest distortion level with a noticeable flicker effect. With the flicker approach, the duration of the subjective test can be reduced by 50% while doubling the perceptual sensitivity compared to the standard binary search approach, in which the compressed images are compared side by side with their original.
Figure 1 shows two original images from the KonJND-1k database, their JPEG and BPG compressed versions corresponding to the first JND points, and SSIM (Wang et al, 2004) maps of the structural similarity between the original and compressed version. From Figure 1, it is visually very difficult to find the differences between the original and its compressed versions. Furthermore, in Figure 1(e) and Figure 1(f), the regions with structural information degradation can be observed. For Figure 1(a), it is the lower part of the original image, while for Figure 1(b), it is the central part of the image. The differences in quality can be seen visually if parts of the images are enlarged, as shown in Figure 2, with 150x150 pixel patches. In Figure 2, a blocking effect is observed with the JPEG compressed image, while a blurring effect is observed with the BPG compressed image. In both cases, the SSIM image similarities are close to the maximum value, and the PSNR-HVS-M values are very high. The PSNR values are greater than 30 dB.
(a) image SRC0336.bmp (900 kB) (b) image SRC0845.bmp (900 kB)
(c) JPEG JND #1 image (QF=62, 32.2 kB) (d) BPG JND #1 image (QP=29, 19.5 kB)
SSIM=0.98 PSNR=30.09 dB PSNR-HVS-M=41.74 dB
SSIM=0.98 PSNR=34.19 dB PSNR-HVS-M=39.98 dB
(e)
(f)
Figure 1 - (a) (b) original images, (c) JPEG JND #1 compressed image, (d) BPG JND #1 compressed image, (e) structural similarity map for JPEG image, and (f) structural similarity map for BPG image
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp. 1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
(c)
(d)
Figure 2 - Patches of: (a) original image SRC0336.bmp, (b) JPEG compressed image, (c) original image SRC0845.bmp and (d) BPG compressed image
Subjective tests conducted by (Alakuijala et al, 2020) showed that H.265/HEVC Intra and JPEG 2000 codecs, with a noticeable blurring effect, generally result in better rate distortion curves when compared to JPEG, with a dominant blocking effect. This means that by blurred (H.265/HEVC Intra and JPEG 2000) images, more visually pleasing images can be obtained.
The content of the source images of the KonJND-1k database is wide, including the categories of people, animals, plants, things, buildings and transport means. Spatial information (SI) and colorfulness (CF) are used as indicators of the complexity of images that need to be compressed. While the definition of colorfulness is unified, different approaches are used to determine spatial information. After determining the local values of image gradient magnitudes, the spatial information can be determined as their maximum value, mean value, root mean square value or as standard
deviation of local scores. Although the ITU recommends using the maximum value, (Yu & Winkler, 2013) has showed that it is better to use the mean gradient magnitude (SImean) to estimate the complexity of the image in compression. Thus, Figure 3 shows scatterplots of SI mean versus CF, along with the convex hulls for two subsets of original KonJND-1k database images. From Figure 3, it can be concluded that both subsets have approximately the same dynamic range of both indicators and that the contents are almost uniformly distributed within their convex hulls.
Figure 3 - Spatial information (SImean) versus colorfulness (CF) plots of the source images belonging to the: (a) JPEG and (b) BPG subsets
Visually lossless compression and objective quality assessment
Objective image quality assessment measures are often used with the aim of achieving the desired quality of the visually lossless compressed image. In some papers, visually lossless quality is defined using fixed thresholds of objective measures, and algorithms for predicting objective values on the boundary between visually lossless and visually lossy compression can also be used. Therefore, it is interesting to analyze the relationship between the objective quality scores and the parameters that control the compression, together with the positions of this boundary.
Figure 4 shows the dependences of PSNR, PSNR-HVS-M and bpp on the quality factor of JPEG compression (rate distortion curves, RDCs). A subset of the KonJND-1k images used for JPEG compression is utilized. Additionally, the positions of the first JND points obtained in the subjective tests (marked with x symbols) are shown.
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
Figure 4 - RDCs PSNR, PSNR-HVS-M andbpp on QF, along with JND #1 points, fora subset of scene images taken from the KonJND-1k database and compressed by JPEG
From Figure 4, no relationship can be observed between the objective quality scores (PSNR, PSNR-HVS-M and bpp) and the QF values of the first JND points (visually lossless compression). It can be observed that the dynamic range of PSNR is significantly higher than the dynamic range of PSNR-HVS-M quality scores. Furthermore, some PSNR and PSNR-HVS-M rate distortion curves show a non-monotonous dependence on the quality factor, known as the strange effect (Bondzulic et al, 2022). This effect does not exist in bpp curves. Similarly, Figure 5 shows the RDCs, along with the positions of the first JND points on the subset of the KonJND-1k images that were used for BPG compression. In this case, the RDCs are monotonous so that the strange effect does not exist for BPG compressed images.
Figure 5 - RDCs PSNR, PSNR-HVS-M and bpp on QP, along with JND #1 points, for a subset of scene images taken from the KonJND-1k database and compressed by BPG
Table 1 provides the boundary values of QF/QP, bpp and objective quality scores for the first JND points of JPEG and BPG subsets. The positions of the first JND points are found in a wide range of PSNR quality values, of almost 20 dB for the JPEG subset, and about 15 dB for the BPG subset. According to PSNR-HVS-M, the range corresponding to the positions of the first JND points on both subsets is about 14 dB. If bpp is used as an indicator, it can be concluded that the positions of the first JND points can be reached by applying lower bpp values in BPG compressed images, which means that a higher degree of compression is achieved than in JPEG compressed images. Additionally, a wide range of QF factors is observed in JPEG compressed images (from 31 to 79). For BPG compressed images, the boundary values of the QP quantization parameter are 25 and 34.
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
Table 1 - Boundary values of PSNR, PSNR-HVS-M, QF/QP and bpp of the first JND points of the JPEG and BPG subsets
Subset JPEG BPG
PSNR minimum 23.3149 28.6247
maximum 42.9831 43.7344
PSNR-HVS-M minimum 37.5680 34.7033
maximum 50.3154 48.9101
QF/QP minimum 31 25
maximum 79 34
bpp minimum 0.1841 0.0237
maximum 2.1429 1.9232
With the histograms of QF/QP values, Figure 6, it can be concluded that the positions of the first JND points of JPEG compressed images are grouped between the QF values of 40 and 45, while in BPG compressed images, the QP values are grouped around 30.
30 35 40 45 50 55 60 65 70 75 80 24 25 26 27 28 29 30 31 32 33 34
QF JND #1 QP JND #1
(a) (b)
Figure 6 - Histograms of: (a) JND #1 JPEG quality factors, and (b) JND #1 BPG quantization parameters for the subsets of the KonJND-1k database
Identification of the features for visually lossless compression
In recent research (Pavlovic et al, 2023), the relationship between simple features derived from the original uncompressed image and the position of the first JND point has been analyzed, with the aim of predicting the values of PSNR JnD #1 or bpp JND #1. Here, the analysis is performed on both subsets of the KonJND-1k database, and for the first time on BPG compressed images. The linear correlation coefficient (LCC),
the root mean square error (RMSE) and the sum of the squared error (SSE) are used in the analysis as quantitative indicators. These values are calculated between the bpp/PSNR JND #1 predictions obtained from the original image features and the ground truth bpp/PSNR values obtained in subjective tests. The second-order mapping law is used to map feature values into JND predictions.
To predict the boundary between visually lossless and visually lossy compression, simple features that have already been used in the literature as complexity measures or can be related to the complexity of the image are chosen. Only one of the adopted features uses color information, while the other features are calculated on grayscale versions of the original images. The following features have been adopted:
• Entropy, E1; it quantifies the information content of the image and can be used to characterize the randomness of the original image intensity distribution.
• Standard Deviation, STD; it is a statistical measure of the pixel intensity values variability.
• Spatial Information, SI, and Spatial Frequency, SF; these are gradient-based features of the image, calculated on the basis of local gradient magnitudes obtained by applying Sobel spatial masks (SI), i.e., differences of pixel intensity values along rows and columns (SF) (Pavlovic et al, 2023). The final values are calculated as their mean value (mean), the root mean square value (rms) and the standard deviation (std).
• Image Power Spectra Slope, IPSS; it is calculated as the slope of the fitted line in the log-log plot of radially averaged spectral power (Pavlovic et al, 2023).
• Two-dimensional Entropy, E2, Contrast, Correlation, Energy and Homogeneity; this feature set is calculated based on the gray level co-occurrence matrix (GLCM), known in image texture analysis and classification (Corchs et al, 2016). The frequencies of occurrence of pairs of gray levels at the positions (m,n) and (m+1,n+1) are used to determine the GLCM.
• Compression Ratio, CRf; as a feature, here, it is calculated as the ratio of the size of the uncompressed image to the size of its JPEG compressed version with a quality factor of QF=100 (Corchs et al, 2016).
• Edge Density, ED; represents the percentage of edge pixels in relation to the dimensions of the image. The well-known Canny detector is used to determine edge pixels.
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
• Shape, a, and scaling, a, parameters of the Generalized Gaussian Distribution (GGD), which models the distribution of Mean Subtracted Contrast Normalized (MSCN) coefficients. The MSCN coefficients distribution parameters are used in the literature to determine the type of degradation and image quality assessment (Mittal et al, 2012).
• Mean Percentage Intensity Uniformity, mPIU; it is determined based on the uniformity within 8x8 non-overlapping blocks of the original image (Goerner et al, 2013).
• Colorfulness, CF; can be used as an estimator of the variety and intensity of colors in an image. It is calculated based on the mean values and standard deviations within the planes obtained after the conversion of the RGB color image into the opponent color space (Hasler & Suesstrunk, 2003).
Tables 2 and 3 show the degree of agreement between the predictions and the ground truth PSNR and bpp values for both subsets of the KonJND-1k database.
Table 2 - Agreement between the predicted and ground truth PSNR (PSNR JND #1) for the KonJND-1k image subsets
PSNR JND #1
Subset JPEG BPG
LCC RMSE SSE LCC RMSE SSE
Entropy, E1 0.3943 3.049 4658 0.4663 2.389 2859
Standard Deviation, STD 0.3043 3.167 5024 0.2165 2.636 3482
Spatial Information, SImean 0.9056 1.407 992.3 0.8585 1.385 960.7
Spatial Information, SIrms 0.8628 1.681 1415 0.7268 1.855 1723
Spatial Information, SIstd 0.7807 2.077 2162 0.6065 2.147 2309
Spatial Frequency, SFmean 0.9178 1.32 873 0.8662 1.349 912.3
Spatial Frequency, SFrms 0.9042 1.42 1010 0.7634 1.744 1524
Spatial Frequency, SFstd 0.8545 1.727 1495 0.6727 1.998 2000
Image Power Spectra Slope, PSS 0.7209 2.304 2660 0.5568 2.243 2521
Entropy, E2 0.7353 2.253 2543 0.7475 1.794 1612
Contrast 0.8438 1.784 1595 0.7059 1.913 1833
Correlation 0.6867 2.412 2914 0.5763 2.207 2440
Energy 0.1043 3.306 5477 0.2319 2.627 3457
PSNR JND #1
Subset JPEG BPG
LCC RMSE SSE LCC RMSE SSE
Homogeneity 0.7446 2.219 2468 0.7878 1.663 1385
Compression Ratio, CRf 0.8246 1.881 1772 0.8508 1.419 1009
Edge Density, ED 0.7921 2.029 2063 0.8294 1.509 1140
GGD, shape а 0.4574 2.956 4378 0.4883 2.357 2782
GGD, scale а 0.7706 2.119 2249 0.7597 1.756 1545
mPIU 0.7253 2.289 2624 0.5818 2.196 2416
Colorfulness, CF 0.3036 3.167 5026 0.2605 2.607 3405
Table 3 - Agreement between the predicted and ground truth bpp (bpp JND #1) for the
KonJND-1k image subsets
bpp JND #1
Subset JPEG BPG
LCC RMSE SSE LCC RMSE SSE
Entropy, E1 0.4620 0.2758 38.12 0.4422 0.2686 36.14
Standard Deviation, STD 0.2798 0.2986 44.67 0.2178 0.2923 42.79
Spatial Information, SImean 0.9052 0.1322 8.759 0.9052 0.1273 8.116
Spatial Information, SIrms 0.8255 0.1755 15.43 0.7934 0.1823 16.65
Spatial Information, SIstd 0.6974 0.2229 24.89 0.6439 0.2291 26.3
Spatial Frequency, SFmean 0.9006 0.1352 9.153 0.9302 0.1099 6.054
Spatial Frequency, SFrms 0.8500 0.1638 13.44 0.8534 0.1561 12.21
Spatial Frequency, SFstd 0.7627 0.2012 20.27 0.7459 0.1994 19.93
Image Power Spectra Slope, PSS 0.6291 0.2418 29.28 0.6131 0.2366 28.04
Entropy, E2 0.7926 0.1897 18.02 0.7824 0.1865 17.43
Contrast 0.8275 0.1746 15.28 0.8218 0.1706 14.58
Correlation 0.6753 0.2294 26.36 0.6859 0.2179 23.79
Energy 0.1703 0.3065 47.06 0.2034 0.2932 43.07
Homogeneity 0.7861 0.1922 18.51 0.7854 0.1854 17.22
Compression Ratio, CRf 0.8131 0.181 16.42 0.8445 0.1604 12.89
Edge Density, ED 0.8061 0.1841 16.97 0.7975 0.1807 16.35
GGD, shape а 0.5479 0.2602 33.91 0.4924 0.2606 34.03
GGD, scale а 0.7253 0.2141 22.97 0.7562 0.1959 19.23
mPIU 0.7670 0.1996 19.95 0.7415 0.2009 20.23
Colorfulness, CF 0.3430 0.2921 42.76 0.1247 0.2971 44.23
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
It can be concluded that gradient-based spatial information (SI) and spatial frequency (SF) are the two best features for prediction, with SF performing better. Additionally, the earlier observation that the mean gradient magnitude is a better predictor than the root mean square value or the standard deviation is confirmed (see SImean/SFmean vs. SIrms/SFrms and SIstd/SFstd). If LCC is considered as a quantitative indicator, it can be concluded that the performance of these two most reliable prediction features is worse in predicting PSNR JND #1 within the BPG subset, and in contrast, the success in predicting bpp JND #1 is observed within the BPG subset.
Also, in earlier research related to JPEG compressed images, it has been concluded that it is better to predict PSNR than bpp, which is also confirmed here. However, it can be concluded here that, for BPG compressed images, it is better to use bpp predictions than PSNR. By predicting bpp based on SFmean, the degree of agreement with the ground truth values is better than 90% for both subsets. Apart from the mentioned two features, contrast, edge density and a compression ratio stand out as good predictors.
As spatial frequency proved to be the best predictor, Figure 7 presents the relationship between SFmean and the ground truth positions of JND #1 points. On both subsets, a linear relationship can be observed between the feature values and the ground truth bpp JND #1 positions, while the trends slightly deviate from the linear law at PSNR JND #1 points. In this way, it is confirmed that the second-order mapping law is a good choice in predicting the positions of JND #1 points.
Additionally, it can be observed that the positions of the first JND points can be reached with lower bpp values, i.e., with a higher degree of compression using BPG. In addition to a higher degree of compression, the PSNR objective quality of BPG compressed images is also better. Finally, it can be concluded that BPG compression at the JND threshold provides both a higher degree of compression and a better PSNR objective quality than JPEG compression.
20
SF
mean
(a)
20 SF
mean
(b)
Figure 7 - Spatial frequency and: (a) PSNR and (b) bpp of the first JND points on the
KonJND-1k image subsets
The optimal functions for mapping the SFmean feature into PSNR JND #1 and bpp JND #1 predictions determined on the KonJND-1k image database results are given in Table 4. In this case, the second-order laws used to map SFmean to PSNR JND #1 must be decreasing functions (see Fig. 7(a)) which remain constant after reaching the minimum values (25.4 dB with JPEG and 29.5 dB with BPG compression). Without this restriction, they would be U-shaped symmetric curves (parabolas). From the law of mapping SFmean to bpp JND #1 of JPEG compressed images, it can be concluded that when SFmean tends to zero, bpp JND #1 tends to 0.2504. For BPG compressed images, it is necessary to introduce a prediction limit (SF mean=0. 75) because without it bpp JND #1 values would be negative as SFmean approaches zero (see Fig. 7(b)).
However, if the relationship between SFmean and PSNR-HVS-M of the first JND points is analyzed, Figure 8, it can be concluded that the quality of JPEG compressed images is better. This quality inversion can be explained by the fact that PSNR is a simple measure that uses the difference of the amplitudes between individual pixels in the quality assessment, while PSNR-HVS-M additionally takes into account the characteristics of the human visual system. Thus, there is a decorrelation between SFmean and PSNR-HVS-M values, so that instead of an orderly trend of scores there is a cloud of points in the 2D space SFmean versus PSNR-HVS-M.
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
Table 4 - Optimal functions for mapping the SFmean feature into PSNR JND #1 and bpp JND #1 predictions determined on the KonJND-1k subsets
JPEG
PSNR PSNR() fo.01376SFl„ - 0.8645SFmean + 38.97,SFmea„ < 31 PSmSF") '1 25.4, SF.a.> 31
bpp bpp (SFmean)' -0.0003719SF;Kan + 0.05558SF„,mn + 0.2504
BPG
PSNR (0.01336SF2 -0.7618SF + 40.36,SF < 28 PSNR(SF-) '1 29.5, SFmean > 28
bpp bpp() J-0.0003425SF1, + 0.05989SFme„n - 0.02376,SFmean > 0.75 ЬРР(SF~) '1 0.0201, SFme^n ^ 0.75
Figure 8 - Spatial frequency and PSNR-FIVS-M of the first JND points on the KonJND-1k
image subsets
JPEG and BPG visually lossless compression comparison
In several previous illustrations, it has been observed that, at the JND threshold, a higher degree of compression can be obtained if BPG compression is used instead of JPEG. This observation is further illustrated in Figure 9, where the relationship between the achieved gain (compression ratio, Gf) calculated as the ratio of the size of the
uncompressed image and its visually lossless version and the SI/SF feature is shown. These scatter plots confirm the previous observation, whereby it can be additionally concluded that the degree of BPG compression is significantly higher than JPEG for lower values of SI/SF, i.e., with low complexity images (with large homogeneous and without textured regions). The degree of compression in both cases is approximately the same for images with higher SI/SF values (images rich in details). On the average, Table 5, the degree of compression at the JND threshold is two times higher if BPG is used instead of JPEG (80.5 vs. 39.4). From Table 5, in which the boundary values of compression are given, it can be concluded that the BPG compression ratio Gf goes over 1000 times (in Figure 9 it is not visible because the dynamic range is shown up to a value of 300) in images with uniform regions. The maximum compression ratio using JPEG is approximately 130. From Figure 9, it can be additionally concluded that SI and SF can be used for reliable prediction of the achieved gain. In this case, due to the large dynamic range and the slope of the dependence of Gf on the mentioned characteristics, it is desirable to use the mapping law of a higher order.
300
250
^ 200 О
.s 150
cd
100
50
0
0
40
80
SI
m
(a)
300
250
200
150
100
50
0
0
20
SF
m
(b)
Figure 9 - The gain achieved by applying visually lossless compression versus: (a) spatial information and (b) spatial frequency
Table 5 - Gain boundary values of visually lossless compression for the KonJND-1k
subsets
Subset JPEG BPG
gain, Gf minimum 11.1996 12.4795
maximum 130.3536 1012.7
mean 39.3763 80.5105
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
Conclusion
The conducted analysis on visually lossless image compression using the KonJND-1k database confirmed some earlier observations about the prediction of the first JND point position of JPEG compressed images, but also brought new conclusions about the prediction of the first JND position of BPG compressed images.
It has been confirmed that the prediction of the first JND point position, as the boundary between visually lossless and visually lossy compression, in JPEG compression can be achieved using simple features derived from the original uncompressed image. The gradient-based features spatial information and spatial frequency are singled out as reliable features for prediction, where the degree of agreement with the results of subjective tests is greater than 90% for both indicators, PSNR and bits per pixel.
The analysis conducted on the prediction of the first JND point position of BPG compressed images has showed that spatial information and spatial frequency can also be used for reliable prediction. The degree of agreement between the predictions derived from these features and the results of subjective tests is greater than 90% if the representation is predicted in bits per pixel. If the PSNR of the first JND point is predicted, the degree of agreement is worse and is around 86%. In addition to these two features, good predictors are contrast, edge density and compression ratio.
A special contribution of the paper is the comparative analysis of the compression ratio on the boundary between visually lossless and visually lossy JPEG and BPG compression. The analysis has showed that the advantage in the degree of compression is on the side of BPG coding, where on average the degree of compression is twice as high compared to JPEG (80 versus 40). The degree of compression of BPG is significantly higher than JPEG for low complexity images (with homogeneous regions), slightly higher for medium complexity images, while for high complexity images with pronounced texture, the degree of compression is approximately the same for both codecs.
Due to the increasing use of images from the infrared part of the spectrum, visually lossless compression in this spectral band will be analyzed in further work. The goal is to conduct subjective tests and additional analyzes using objective image quality assessment measures.
References
Alakuijala, J., Boukortt, S., Ebrahimi, T., Kliuchnikov, E., Sneyers, J., Upenik, E., Vandevenne, L., Versari, L. & Wassenberg, J. 2020. Benchmarking JpEg XL image compression. In: Schelkens, P. & Kozacki, T. (Eds.) Proceedings SPIE Photonics Europe: Optics, Photonics and Digital Technologies for Imaging Applications VI, Online Only, France, 11353, pp.113530X-1-113530X-20, April 06-10. Available at: https://doi.org/10.1117/12.2556264.
Bellard, F. 2018. BPG Image format. Bellard.org, 21 April [online]. Available at: https://bellard.org/bpg/ [Accessed: 05. April 2024].
Bull, D. & Zhang, F. 2021. Intelligent Image and Video Compression: Communicating Pictures, 2nd Edition. Academic Press. ISBN: 9780128203545.
Bondzulic, B., Stojanovic, N., Petrovic, V., Pavlovic, B. & Milicevic, Z. 2021. Efficient Prediction of the First Just Noticeable Difference Point for JPEG Compressed Images. Acta Polytechnica Hungarica, 18(8), pp.201-220. Available at: https://doi.org/10.12700/APH.18.8.2021.8.11.
Bondzulic, B., Bujakovic, D., Li, F. & Lukin, V. 2022. On strange images with application to lossy image compression. Radioelectronic and Computer Systems, 104(4), pp.143-152. Available at: https://doi.org/10.32620/reks.2022A11.
Bondzulic, B., Pavlovic, B., Stojanovic, N., Petrovic, V. & Bujakovic, D. 2024. A simple and reliable approach to providing a visually lossless image compression. The Visual Computer, International Journal of Computer Graphics, 40, pp.3747-3763. Available at: https://doi.org/10.1007/s00371-023-03062-y.
Cai, Y.-Q., Zou, H.-X. & Yuan, F. 2019. Adaptive compression method for underwater images based on perceived quality estimation. Frontiers of Information Technology & Electronic Engineering, 20(5), pp.716-730. Available at: https://doi.org/10.1631/FITEE.1700737.
Corchs, S.E., Ciocca, G., Bricolo, E. & Gasparini, F. 2016. Predicting complexity perception of real world images. PLoS ONE, 11(6), e0157986. Available at: https://doi.org/10.1371/journal.pone.0157986.
Fan, C., Lin, H., Hosu, V., Zhang, Y., Jiang, Q., Hamzaoui, R. & Saupe, D. 2019. SUR-Net: Predicting the satisfied user ratio curve for image compression with deep learning. In: 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), Berlin, Germany, pp.1-6, June 5-7. Available at: https://doi.org/10.1109/QoMEX.2019.8743204.
Fiorucci, F., Baruffa, G. & Frescura, F. 2012. Objective and subjective quality assessment between JPEG XR with overlap and JpEG 2000. Journal of Visual Communication and Image Representation, 23(6), pp.835-844. Available at: https://doi.org/10.1016/jJvcir.2012.04.011.
Goerner, F.L., Duong, T., Stafford, R.J. & Clarke, G.D. 2013. A comparison of five standard methods for evaluating image intensity uniformity in partially parallel imaging MRI. Medical Physics, 40(8), art.number:082302. Available at: https://doi.org/10.1118/14816306.
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
Hasler, D. & Suesstrunk, S.E. 2003. Measuring colourfulness in natural images. In: Rogowitz, B.E. & Pappas, T.N. (Eds.) Proceedings Electronic Imaging: Human Vision and Electronic Imaging VIII, Santa Clara, CA, USA, 5007, pp.8795, January 20-24. Available at: https://doi.org/10.1117/12.477378.
Jin, L., Lin, J.Y., Hu, S., Wang, H., Wang, P., Katsavounidis, I., Aaron, A. & Kuo, C.-C. J. 2016. Statistical study on perceived JPEG image quality via MCL-JCI dataset construction and analysis. In: IS&T International Symposium on Electronic Imaging - Image Quality and System Performance XIII, San Francisco, CA, USA, art.ID:art00026, February 14-18. Available at: https://doi.org/10.2352/ISSN.2470-1173.2016.13.IQSP-222.
Kovalenko, B., Lukin, V., Kryvenko, S., Naumenko, V. & Vozel, B. 2022. Prediction of parameters in optimal operation point for BPG-based lossy compression of noisy images. Ukrainian Journal of Remote Sensing, 9(2), pp.412. Available at: https://doi.org/10.36023/ujrs.2022.9.2.212.
Kovalenko, B. & Lukin, V. 2023. Analysis of distortions due to BPG-based lossy compression of noise-free and noisy images. Herald of Khmelnytskyi National University. Technical Sciences, 325(1), pp.128-135.
Krivenko, S.S., Krylova, O., Bataeva, E. & Lukin, V.V. 2018. Smart lossy compression of images based on distortion prediction. Telecommunications and Radio Engineering, 77(17), pp. 1535-1554. Available at:
https://doi.org/10.1615/TelecomRadEng.v77.i17.40.
Li, F., Lukin, V., Okarma, K. & Fu, Y. 2021. Providing a Desired Quality of BPG Compressed Images for FSIM Metric. In: 2021 IEEE 3rd International Conference on Advanced Trends in Information Theory (ATIT), Kyiv, Ukraine, pp.10-14, December 15-17. Available at:
https://doi.org/10.1109/ATIT54053.2021.9678522.
Li, F., Lukin, V., Ieremeiev, O. & Okarma, K. 2022. Quality control for the BPG lossy compression of three-channel remote sensing images. Remote Sensing, 14(8), art. number: 1824. Available at:
https://doi.org/10.3390/rs14081824.
Li, F., Abramov, S., Dohtiev, I. & Lukin, V. 2024. Advantages and drawbacks of two-step approach to providing desired parameters in lossy image compression. Advanced Information Systems, 8(1), pp.57-63. Available at: https://doi.org/10.20998/2522-9052.2024.1.07.
Lin, H., Hosu, V., Fan, C., Zhang, Y., Mu, Y., Hamzaoui, R. & Saupe, D. 2020. SUR-FeatNet: Predicting the satisfied user ratio curve for image compression with deep feature learning. Quality and User Experience, 5, art.number:5. Available at: https://doi.org/10.1007/s41233-020-00034-1.
Lin, H., Chen, G., Jenadeleh, M., Hosu, V., Reips, U.-D., Hamzaoui, R. & Saupe, D. 2022. Large-Scale Crowdsourced Subjective Assessment of Picturewise Just Noticeable Difference. IEEE Transactions on Circuits and Systems for Video Technology, 32(9), pp.5859-5873. Available at: https://doi.org/10.1109/TCSVT.2022.3163860.
Liu, X., Chen, Z., Wang, X., Jiang, J. & Kwong, S. 2018. JND-Pano: Database for just noticeable difference of JPEG compressed panoramic images. In: Hong, R., Cheng, W.H., Yamasaki, T., Wang, M. & Ngo, C.W. (Eds.) Advances in Multimedia Information Processing. PCM 2018. Lecture Notes in Computer Science, 11164, pp.458-468. Cham: Springer. Available at:
https://doi.org/10.1007/978-3-030-00776-8_42.
Liu, H., Zhang, Y., Zhang, H., Fan, C., Kwong, S., Kuo C.-C. J. & Fan, X. 2020. Deep Learning-Based Picture-Wise Just Noticeable Distortion Prediction Model for Image Compression. IEEE Transactions on Image Processing, 29, pp.641-656. Available at: https://doi.org/10.1109/TIP.2019.2933743.
Liu, Y., Jin, J., Xue, Y. & Lin, W. 2023. The First Comprehensive Dataset with Multiple Distortion Types for Visual Just-Noticeable Differences. In: 2023 IEEE International Conference on Image Processing (ICIP), Kuala Lumpur, Malaysia, pp.2820-2824, October 08-11. Available at: https://doi.org/10.1109/ICIP49359.2023.10221977.
Mikhailiuk, A., Ye, N. & Mantiuk, R.K. 2021. The effect of display brightness and viewing distance: A dataset for visually lossless image compression. In: IS&T International Symposium on Electronic Imaging: Human Vision and Electronic Imaging, Virtual Event, 33, art.ID:art00005, January 21st. Available at: https://doi.org/10.2352/ISSN.2470-1173.2021.11.HVEI-152.
Mittal, A., Moorty, A.K. & Bovik, A.C. 2012. No-Reference Image Quality Assessment in the Spatial Domain. IEEE Transactions on Image Processing, 21(12), pp.4695-4708. Available at: https://doi.org/10.1109/TIP.2012.2214050.
Pavlovic, B., Bondzulic, B., Stojanovic, N., Petrovic, V. & Bujakovic, D. 2023. Prediction of the first just noticeable difference point based on simple image features. In: 2023 Zooming Innovation in Consumer Technologies Conference (ZINC), Novi Sad, Serbia, pp. 125-130, May 29-31. Available at: https://doi.org/10.1109/ZINC58345.2023.10173865.
Ponomarenko, N., Silvestri, F., Egiazarian, K., Carli, M., Astola, J. & Lukin, V. 2007. On between-coefficient contrast masking of DCT basis functions. In: 3rd International Workshop on Video Processing and Quality Metrics for Consumer Electronics (VPQM), Scottsdale, Arizona, USA, pp.1-4, January 25-26 [online]. Available at: http://ponomarenko.info/vpqm07_p.pdf [Accessed: 05. April 2024].
Poth, M., Trpovski, Z. & Loncar-Turukalo, T. 2020. Analysis and improvement of JPEG compression performance using custom quantization and block boundary classifications. Acta Polytechnica Hungarica, 17(6), pp.171-191. Available at: https://doi.org/10.12700/APH.17.6.2020.6.10.
Saha, S. & Vemuri, R. 2000. An analysis on the effect of image features on lossy coding performance. IEEE Signal Processing Letters, 7(5), pp.104-107. Available at: https://doi.org/10.1109/97.841153.
Shen, X., Ni, Z., Yang, W., Zhang, X., Wang, S. & Kwong, S. 2020. A JND Dataset Based on VVC Compressed Images. In: 2020 IEEE International Conference on Multimedia & Expo Workshops (ICMEW), London, United Kingdom, pp.1-6, July 6-10. Available at:
https://doi.org/10.1109/ICMEW46912.2020.9105955.
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
Shen, X., Ni, Z., Yang, W., Zhang, X., Wang, S. & Kwong, S. 2021. Just Noticeable Distortion Profile Inference: A Patch-Level Structural Visibility Learning Approach. IEEE Transactions on Image Processing, 30, pp.26-38. Available at: https://doi.org/10.1109/TIP.2020.3029428.
Testolina, M., Hosu, V., Jenadeleh, M., Lazzarotto, D., Saupe, D. & Ebrahimi T. 2023. JPEG AIC-3 Dataset: Towards Defining the High Quality to Nearly Visually Lossless Quality Range. In: 2023 15th International Conference on Quality of Multimedia Experience (QoMEX), Ghent, Belgium, pp.55-60, June 2022. Available at: https://doi.org/10.1109/QoMEX58391.2023.10178554.
Wang, Z. & Bovik, A.C. 2009. Mean squared error: Love it or leave it? A new look at Signal Fidelity Measures. IEEE Signal Processing Magazine, 26(1), pp.98117. Available at: https://doi.org/10.1109/MSP.2008.930649.
Wang, Z., Bovik, A.C., Sheikh, H.R. & Simoncelli, E.P. 2004. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4), pp.600-612. Available at:
https://doi.org/10.1109/TIP.2003.819861.
Yu, H. & Winkler, S. 2013. Image complexity and spatial information. In: 2013 Fifth International Workshop on Quality of Multimedia Experience (QoMEX), Klagenfurt am Worthersee, Austria, pp. 12-17, July 3-5. Available at: https://doi.org/10.1109/QoMEX.2013.6603194.
Zemliachenko, A., Ponomarenko, N., Lukin, V., Egiazarian, K. & Astola, J. 2016. Still image/video frame lossy compression providing a desired visual quality. Multidimensional Systems and Signal Processing, 27(3), pp.697-718. Available at: https://doi.org/10.1007/s11045-015-0333-8.
Compresion de imagenes JPEG y BPG sin perdida visual a traves de la base de datos KonJND-1k
Boban P. Bondzulica, Nenad M. Stojanovica, autor de correspondencia, Vladimir V. Lukinb, Sergii S. Kryvenkob
а Universidad de Defensa de Belgrado, Academia Militar, Departamento de Telecomunicaciones e Informatica, Belgrado, Republica de Serbia b Universidad Aeroespacial Nacional, Departamento de Tecnologlas de la Information y la Comunicacion, Kharkiv, Ucrania
CAMPO: telecomunicacion
TIPO DE ARTICULO: articulo cientifico original
Resumen:
Introduccion/objetivo: Este articulo presenta los resultados de la investigacion sobre la compresion de imagenes sin perdida visual, que es de particular interes porque logra un alto grado de compresion, mientras que la calidad visual de la imagen no se ve afectada, es decir, los usuarios finales estan muy satisfechos con la calidad de la imagen. El analisis se llevo a cabo utilizando la base de datos KonJND-1k de
imageries a gran escala disponible publicamente, que contiene los resultados de pruebas subjetivas en imagenes comprimidas JPEG y BPG.
Metodos: Gracias a la disponibilidad de imagenes de la base de datos KonJND-1k, se analiza la dependencia de las evaluaciones objetivas de la calidad de la imagen en los parametros que controlan el grado de compresion de las senales de origen (factorde calidad para JPEG y parametro de cuantificacion para BPG). Los resultados de las pruebas subjetivas sin perdida visual se utilizan para un analisis profundo de los valores limite y tipicos de los parametros que controlan estos dos tipos de compresion, asi como para el analisis de los valores correspondientes de las puntuaciones de calidad objetivas. Ademas, se han identificado caracteristicas fiables para predecir el limite entre la compresion visual sin perdida y la compresion visual con perdida. Para ello, se utiliza el grado de acuerdo entre las predicciones y los valores de verdad fundamental de la relacion senal-ruido (PSNR) de pico y la representacion de la imagen en bits por pixel (bpp). La relacion de compresion visual sin perdida se utiliza para comparar las tecnicas JPEG y BPG.
Resultados: Se muestra que el limite entre la compresion de imagenes visual sin perdida y visual con perdida se encuentra en un amplio rango de valores de PSNR (aproximadamente 20 dB para JPEG y 15 dB para BPG). Los valores del factor de calidad de compresion de imagenes JPEG correspondientes en este umbral tambien varian ampliamente de 31 a 79, con una concentracion entre 40 y 45. Para el codificador BPG, los valores del parametro de cuantificacion se agrupan alrededor de 30, y los valores limite son 25 y 34. Ademas, se muestra que este limite se puede determinar de forma fiable basandose en caracteristicas simples derivadas de la imagen original sin comprimir. Las caracteristicas basadas en gradientes conocidas como frecuencia espacial e informacion espacial demostraron ser los mejores predictores. El grado de acuerdo entre las predicciones obtenidas a partir de estas caracteristicas con los valores de verdad fundamental de PSNR y bpp en ambos tipos de compresion es superior al 85%. Un analisis comparativo ha demostrado que, utilizando la compresion BPG, es posible, en promedio, lograr una relacion de compresion de compresion sin perdida visual dos veces mayor que para JPEG (80 frente a 40).
Conclusion: Aunque se logra un alto grado de acuerdo entre las predicciones y los valores de verdad fundamental de PSNR y bpp del limite entre la compresion sin perdida visual y la compresion con perdida visual, existe la necesidad de desarrollar nuevos enfoques de prediccion, especialmente con la tecnica BPG, que a traves de la relacion de compresion demostro ser superior a la tecnica JPEG. Las bases de datos existentes utilizadas para el analisis de la compresion sin perdida visual contienen imagenes en color de la parte visible del espectro
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
electromagnetico. Considerando el uso creciente de imageries de la parte infrarroja del espectro, existe la necesidad de realizar pruebas similares en este rango espectral.
Palabras claves: compresion BPG, compresion JPEG, diferencia apenas perceptible (JND), relacion senal-ruido maxima (PSNR), compresion de imagenes sin perdida visual.
Сжатие изображений JPEG и BPG без визуальных потерь с помощью базы данных KonJND-1k
Бобан П. Бонджулича, Ненад М. Стояновича, корреспондент,
Владимир В. Лукинб, Сергей С. Кривенкоб
а Университет обороны в г. Белград, Военная академия, кафедра телекоммуникаций и информатики, г. Белград, Республика Сербия
б Национальный аэрокосмический университет им. М. Е. Жуковского «Харьковский авиационный институт», Кафедра информационнокоммуникационных технологий им. О.А. Зеленского, г. Харьков, Украина
РУБРИКА ГРНТИ: 28.23.00 Искусственный интеллект;
28.23.15 Распознавание образов. Обработка изображений
ВИД СТАТЬИ: оригинальная научная статья Резюме:
Введение/цель: В данной статье представлены результаты исследований по сжатию изображений без визуальных потерь, которое представляет особый интерес, поскольку позволяет достичь высокой степени сжатия, при этом визуальное качество изображения не ухудшается, т.е. конечные пользователи очень довольны качеством изображения. Анализ проводился с
использованием общедоступной крупномасштабной картинной базы данных KonJND-1k, содержащей результаты субъективных тестов изображений, сжатых в форматах JPEG и BPG.
Методы: Благодаря наличию изображений из базы данных KonJND-1k проанализирована зависимость объективных оценок качества изображения от параметров, управляющих степенью сжатия исходных сигналов (коэффициента качества для JPEG и параметра квантования для BPG). Результаты субъективных тестов без визуальных потерь используются для тщательного анализа граничных и типичных значений параметров, управляющих этими двумя типами сжатия, а также для анализа соответствующих значений объективных показателей качества. Кроме того, были идентифицированы надежные функции для прогнозирования границы между сжатием без визуальных потерь и сжатием с визуальными потерями. Для этой цели используется степень
согласия между прогнозируемыми и фактическими значениями пикового отношения сигнал/шум (PSNR) и представления изображения в битах на пиксель (bpp). Степень сжатия без визуальных потерь используется для сравнения методов JPEG и BPG.
Результаты: Показано, что граница между сжатием изображений без визуальных потерь и сжатием изображений с визуальными потерями находится в широком диапазоне значений PSNR (около 20 дБ для JPEG и 15 дБ для BPG). Соответствующие значения коэффициента качества сжатия изображения JPEG при этом пороге также широко варьируются от 31 до 79 с накоплением от 40 до 45. Значения параметра квантования группируются около 30, а граничные значения составляют 25 и 34. Показано, что эту границу можно надежно определить на основе простых признаков, полученных из исходного несжатого изображения. Градиентные характеристики, известные как пространственная частота и пространственная информация, оказались лучшими предикторами. Степень согласия между прогнозируемыми и фактическими значениями PSNR и bpp при обоих типах сжатия превышает 85%. Сравнительный анализ показал, что при использовании BPG сжатия в среднем можно добиться вдвое большей степени сжатия без визуальных потерь, чем при JPEG (80 против 40).
Выводы: Несмотря на достигнутую высокую степень согласия между прогнозируемыми и основными значениями PSNR и bpp на границе между сжатием без визуальных потерь и сжатием с визуальными потерями, существует необходимость в разработке новых подходов к прогнозированию, особенно с использованием метода BPG, который по коэффициенту сжатия превосходит метод JPEG. Современные базы данных, используемые для анализа сжатия без визуальных потерь, содержат цветные изображения видимой части электромагнитного спектра. Учитывая широкое использование изображений инфракрасной части спектра, возникает необходимость проведения аналогичных испытаний и в этом спектральном диапазоне.
Ключевые слова: сжатие BPG, сжатие JPEG, едва заметная разница (JND), пиковое соотношение сигнал/шум (PSNR), сжатие изображения без визуальных потерь.
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp.1214-1241
VOJNOTEHNICKI GLASNIK / MILITARY TECHNICAL COURIER, 2024, Vol. 72, Issue 3
Компресса JPEG и BPG без губитка визуелних информаци]а на примеру базе KonJND-1k
Бобан П. Бонуулиб3, Ненад М. Сто]ановиба, аутор за преписку,
Владимир В. Лукин6, Cepaej С. Кривенко6
а Универзитет одбране у Београду, Во]на академи]а, Катедра телекомуникаци]а и информатике, Београд, Република Срби]а
б Национални ваздухопловни универзитет, Департман информационо-комуникационих технологи]а, Харков, Украина
ОБЛАСТ: телекомуникаци]е
КАТЕГОРИJА (ТИП) ЧЛАНКА: оригинални научни рад Сажетак:
Увод/циж У овом раду представ^ени су резултати истраживаьа компрессе без губитка визуелних информаци'а. Она 'е од посебног знача'а jер се ьоме оствару'е висок степен компрессе, при чему визуелни квалитет слике ни'е нарушен, па су краjьи корисници веома задовоъни. Анализа 'е спроведена коришПеьем обимне, jавно доступне базе KonJND-1k, ко'а садржи резултате субjeктивних тестова на компримованим сликама JPEG и BPG.
Методе: Захва^ууПи доступности слика базе KonJND-1k
анализирана 'е зависност обjeктивних мера процене квалитета слике од параметара ко има се контролише степен компреси е изворних сигнала (фактор квалитета код JPEG, односно параметар квантизаци'е код BPG). Резултати субjeктивних тестова искоришПени су за дета^ни'у анализу граничних и типичних вредности параметара ко има се контролишу ова два типа компрессе, као и за анализу одговара'убих вредности обjeктивних скорова квалитета. Тако^е, извршена 'е идентификаций поузданих обележ'а за предикциу границе измену компреси е без и са губитком визуелних информаци а. У ту сврху коришПен 'е степен слагаьа измену предикци'а и тачних вредности вршног односа сигнал/шум (PSNR) и репрезентацие слике у битима по пикселу (bpp). Степен компреси'е остварен применом компреси'е без губитка визуелних информаци'а искоришПен 'е за поре^еье техника JPEG и BPG.
Резултати: Показано 'е да се граница измену компреси'е без и са губитком визуелних информаци а налази у широком опсегу вредности PsNr (око 20 dB код JpEg и 15 dB код BPG). Одговара'уПе вредности фактора квалитета слика JPEG на ово граници се, тако^е, налазе у широком опсегу од 31 до 79, са груписаьем измену 40 и 45. Вредности параметра квантизаци'е групишу се око 30, а граничне вредности су 25 и 34. Тако^е, потвр^ено jе да се ова граница може поуздано одредити на основу jедноставних обележ'а изведених из оригиналне некомпримоване слике. Показало се да су
на)бо^и предиктори градуентна обележ1'а позната као просторна фреквенцуа и просторна информацуа. Степен слагала предикцуа добуених из ових обележ]а са тачним вредностима PSNR и bpp код оба типа компрессе веПи )е од 85%. Компаративном анализом доказано )е да се применом компрессе BPG, у просеку, може остварити дупло веПи степен компрессе без губитка визуелних информацуа него применом компрессе JPEG (80 насупрот 40). Закъучак: Иако )е остварен висок степен слагала измену предикцуа и тачних вредности PSNR и bpp на граници измену компрессе без и са губитком визуелних информацуа, постов потреба за разво ем нових приступа предикци е, нарочито код технике BPG ко а се кроз степен компреси е показала супериорном у односу на технику JPEG. Посто]еПе базе ко)е се користе за анализу компреси е без губитка визуелних информаци а су са сликама из видъивог дела електромагнетног спектра. Има]уПи у виду све веПу употребу слика из инфрацрвеног дела спектра, постов потреба за спрово^еъем сличних тестова у овом спектралном опсегу.
К^учне речи: компресса BPG, компресса JPEG, )едва уочъиве разлике (JND), вршни однос сигнал/шум (PSNR), компресса без губитка визуелних информацуа.
EDITORIAL NOTE: The first author of this article, Boban P. Bondzulic, is a current member of the Editorial Board of the Military Technical Courier. Therefore, the Editorial Team has ensured that the double blind reviewing process was even more transparent and more rigorous. The Team made additional effort to maintain the integrity of the review and to minimize any bias by having another associate editor handle the review procedure independently of the editor - author in a completely transparent process. The Editorial Team has taken special care that the referee did not recognize the author's identity, thus avoiding the conflict of interest.
Paper received on: 07.04.2024.
Manuscript corrections submitted on: 24.09.2024.
Paper accepted for publishing on: 25.09.2024.
© 2024 The Authors. Published by Vojnotehnicki glasnik / Military Technical Courier (www.vtg.mod.gov.rs, втг.мо.упр.срб). This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/rs/).
Bondzulic, B. et al, JPEG and BPG visually lossless image compression via KonJND-1k database, pp. 1214-1241
