Objective Image Quality Evaluation for JPEG, JPEG 2000, and Vidware Vision TM

Transcription

1 38 Chung-Hao Chen, Yi Yao, David L. Page, Besma Abidi, Andreas Koschan, and Mongi Abidi, "Objective Image Quality Evaluation for JPEG, JPEG, and Vidware Vision." Advances in Visual Computing, Proceedings of Second International Symposium, ISVC 6, Lake Tahoe, NV, November 6. Objective Image Quality Evaluation for JPEG, JPEG, and Vidware Vision TM Chung-Hao Chen, Yi Yao, David L. Page, Besma Abidi, Andreas Koschan, and Mongi Abidi The University of Tennessee Imaging, Robotics and Intelligent System Laboratory Knoville, Tennessee , USA {cchen1, yyao1, dpage, besma, akoschan, Abstract. In this paper, three compression methods, JPEG, JPEG, and Vidware Vision TM are evaluated by different full- and no-reference objective image quality measures including Peak-Signal-to-Noise-Ratio (PSNR), structural similarity (SSIM), and Tenengrad. In the meantime, we also propose an image sharpness measure, non-separable rational function based Tenengrad (NSRT ), to address whether the compression method is appropriate to be used in a machine recognition application. Based on our eperimental results Vidware Vision TM is more robust to changes in compression ratio and presents gradually degraded performance at a considerably slower speed thus outperforming JPEG and JPEG when the compression ratio is smaller than.7%. Furthermore, the effectiveness of our proposed measurement, NSRT, is also validated via eperiments and performance comparisons with other objective image quality measures. Keywords: JPEG, JPEG, Vidware Vision TM, image quality, mage sharpness, and image compression. 1 Introduction Image compression [1], [], [3] is an essential component of image retrieval [5], recognition [6], and internet applications [7]. The goal of compression is to minimize the size of the data being broadcast or stored, thus minimizing transmission time and storage space while still maintaining a desired quality compared to the original image. One of the most popular lossy compression approaches for still images is JPEG. JPEG [] stands for Joint Photographic Eperts Group, the name of the committee that developed the standard. JPEG compression divides the input image into 8 by 8 piel blocks and calculates the discrete cosine transform (DCT) of each block. A quantizer rounds off these DCT coefficients according to a pre-defined quantization matri. This quantization step produces the "lossy" nature of JPEG. Afterwards, JPEG applies a variable length code to these quantized coefficients. JPEG [1] is a wavelet-based image compression standard that was also created by the Joint Photographic Eperts Group with intention to outperform their L.-W. Chang, W.-N. Lie, and R. Chiang (Eds.): PSIVT 6, LNCS 4319, pp , 6. Springer-Verlag Berlin Heidelberg 6

2 75 C.-H. Chen et al. original JPEG standard. JPEG is based on the idea that coefficients of a transform which decorrelates piels of an image could be coded more effectively than the original piels. If functions of the transform, which is the wavelets transform in JPEG, translate most of the important visual information into a small number of coefficients then the remaining coefficients could be quantized/normalized coarsely or even truncated to zero without introducing noticeable distortions. Compared with JPEG, JPEG [11] not only improves the quality of decompressed images but also supports much lower compression ratios. Recently, Vidware Incorporated developed a product line called Vidware Vision TM [3], [4]. It is divided into three separate categories: Still Image (as a replacement for JPEG), Full Frame video (as a replacement for M-JPEG), and Full Motion video (an H.64 compliant CODEC). Vidware Vision TM Still Image is an integer-based encoding algorithm and varies the size of macro-blocks according to the shape and contents of the image, unlike JPEG where the block size is fied. Therefore, this variable macro-blocking produces reduced blocky artifacts and increases image fidelity. In terms of full-reference image quality measure, PSNR is the most widely used full-reference quality measure. It is appealing because of its simple computation and clear physical meaning [9], [1]. Nevertheless, it is not closely matched to perceived visual quality [8], [9], [1]. Therefore, SSIM [1] is selected for its improved representation of visual perception. The SSIM measure compares local patterns of piel intensities that have been normalized and hence are invariant to luminance and contrast. This method could be seen as complementary to the traditional PSNR approach. A successful recognition requires sufficient local details to differentiate various patterns and the ability to preserve these local details. These details translate into the sharpness of the decompressed images and interpreted by image sharpness measures. Sharpness measures have been traditionally divided into 5 categories [13]: gradient based, variance based, correlation based, histogram based, and frequency domain based methods. Image noise level and artifacts, such as blocking effects, vary with respect to different compression methods as well as various compression ratios. Conventional sharpness measures are not feasible because they are unable to differentiate variations caused by actual image edges from those induced by image noise and artifacts. To avoid artificially elevated sharpness values due to image noise and artifacts, NSRT is proposed as an adaptive measure. The contributions of this paper are: (1) Evaluation of three compression methods, JPEG, JPEG, and Vidware Vision TM by different full- and no-reference objective image quality measures; () The design of NSRT as an adaptive measure, and its ability to evaluate the sharpness of decompressed images for machine recognition applications. The remainder of this paper is organized as follow. Section describes our proposed adaptive sharpness measure, NSRT. Eperimental results are demonstrated in section 3, and Section 4 concludes the paper.

3 Objective Image Quality Evaluation for JPEG, JPEG, and Vidware Vision TM 753 Sharpness Quality Measure Sharpness measures are traditionally used to quantify out-of-focus blur. Nevertheless, their etension to evaluate the sharpness of compressed images is non-trivial. To avoid artificially elevated sharpness values due to image noise and artifacts, adaptive sharpness measures assign different weights to piel gradients according to their local activities. For piels in smooth areas, small weights are used. For piels adjacent to strong edges, large weights are allocated. Adaptive sharpness measures are comprised of two determinant factors: the definition of local activities and the selection of weight functions. Figure 1 depicts the flow chart to compute adaptive sharpness measures. Image gradients Weight Functions Separable Polynomial Non-separable Rational Gradient based sharpness measures Fig. 1. Flow chart to compute adaptive sharpness measures Based on how local activities are described, adaptive sharpness measures can be divided into two groups: separable and non-separable. Separable measures only focus on horizontal and vertical edges. A horizontal image gradient g ( and a vertical image gradient g y ( are computed independently. For instance, g( = f ( + 1, f ( 1, (1) gy( = f ( y+ 1) f ( y 1) In contrast, non-separable measures include the contributions from diagonal edges. An eample of this type of image gradients g ( is given by: g ( = f ( 1, + f ( + 1, f ( y 1) f ( y + 1). () Different forms of weights can be used, among which polynomial and rational functions are two popular choices. The polynomial, to be more specific cubic, and rational functions are also eploited in adaptive unsharp masking [14], [15]. The polynomial weights suppress small variation mostly introduced by image noise and have been proved efficient in evaluating the sharpness of high magnification images [16]. The rational weights emphasize a particular range of image gradients. Taking the non-separable image gradient g ( for eample, the polynomial weights are:

4 754 C.-H. Chen et al. pω ω(, = g( (3) where p ω is a power inde determining the degree of noise suppression. The rational weights can be written as ( k + k ) 1 g( ω ( = (4) g ( + k g( + k where k and k 1 are coefficients associated with the peak position L and width L of the corresponding function, respectively, and comply with the following equations k k = L + 8k k + 1k L 1 1 = Figure illustrates the comparison of different forms of weight functions. 1 (5) ω ().4 L =1, L=35 L =1, L=5. p = ω p =6 ω No weights Fig.. Illustration of weight functions The newly designed weights are then applied to gradient based sharpness measures to construct adaptive sharpness measures. For the Tenengrad measure for instance, the resulting separable measure is given by: [ ( g ( ω ( g ( y ] S = ω + ) (6) M N where ω ( / ω y ( denotes the weights obtained from the horizontal/vertical gradients g ( / g y (. For non-separable methods, the corresponding adaptive Tenengrad is formulated as y [ g ( g ( y ] S = ω ( + ) (7) M N To account for blocking artifacts, we follow the method Wang et al. proposed in [1]. The above measures are divided into two groups: piels within a block and piels at block borders. Accordingly, we compute two values y y

5 Objective Image Quality Evaluation for JPEG, JPEG, and Vidware Vision TM 755 S = M 1 N 1 =, y = mod( 8) mod( 8) ω( [ g ( + g ( ] y B = M =, N 1 y= ω(88 and [ g (88 + g (88 ] y (8) The final image quality measure combines these two values: p Q = S B 1 (9) The adjustable variables in the rational weight based measures include the peak position L, response width L, and weight power p. These parameters can be selected according to the visual perception of the images. In our implementation, piel gradients g ( / g y ( / g( are scaled by their mean before computing the weights. The chosen measure NSRT then has a peak position k of 1 and a weight power p of. The coefficient k 1 is forced to be zero, resulting in a response width of L = 3 L = 35. In practice, different forms of image gradients are employed for deriving weights, as in equations (3) and (4), and computing sharpness, as in equations (6) and (7), for improved robustness to image noise and artifacts. For weight computation the image gradients used are listed in equations (1) and (), while for sharpness computation the image gradients based on the Sobel filters are recommended. 3 Eperimental Results We compare the performance of JPEG, JPEG, and Vidware Vision TM based on a data set composed of compressed images at various compression ratios. The 48 raw color images (4 bits/piel RGB) in the data set include 1 selected images with different resolutions and 36 long range face images (resolution: 7 48). The 1 selected images, as shown in part in Fig. 3 (a)-(d), can be further divided into 4 categories. The first category includes some well-known test images such as the Baboon, Lena, and Peppers. The second category is comprised of standard test patterns, the IEEE resolution chart (sinepatterns.com), Television Color Test Pattern (high-techproductions.com), and Color Test Template (eecs.berkeley.edu). The third category focuses on images used in face and license plate recognition, two eemplary applications of machine recognition. High resolution images are covered in the last category. The second group, as shown in part in Fig. 3 (e)-(f), is a collection of face images of 6 subjects. A total of 6 images per subject are obtained at various distances

6 756 C.-H. Chen et al. (9.5, 1.4, 11.9, 13.4, 14.6, and 15.9 meters) and with different camera s focal length. The camera s focal length is adjusted so that the subject s face presents similar sizes in all 6 images. Each selected image is compressed by JPEG, JPEG, and Vidware Vision TM with different compression ratios varying from.% to 3%. This produces a total of 34 compressed images for each compression method. In addition, the compression ratio used in this paper indicates: R C = (1) O Where C, R, and O represent compression ratio, file size of the compressed image, and file size of the original image respectively. (a) (b) (c) (d) (e) (f) Fig. 3. Illustration of test images, (a) Lena 51 51; (b) IEEE resolution chart ; (c) License plate ; (d) Under vehicle view 56 19; (e) 9.5 meters face image; (f) 15.9 meters face image 3.1 Full-Reference Quality Measures Figure 4(a) illustrates the computed SSIM measure. We could see that JPEG and Vidware Vision TM outperform JPEG for all tested compression ratios. For compression ratios in the range of 7%~1%, JPEG presents a slightly better performance compared with Vidware Vision TM. However, this performance difference diminishes for other tested compression ratios. Figure 4(b) illustrates the computed PSNR measure. As epected, JPEG outperforms JPEG for all tested compression ratios. However, their PSNR values increase quickly with respect to decreased compression ratios, indicating a rapidly degraded image quality. In contrast, the PSNR value of Vidware Vision TM increases at a considerably slower speed. As a result, Vidware Vision TM eventually outperforms both JPEG and JPEG as the compression ratio goes below.7%.

7 Objective Image Quality Evaluation for JPEG, JPEG, and Vidware Vision TM Structural Similarity (SSIM) JPEG JPEG Vidware Compression ratio (%) (a) Peak-Signal-to-Noise-Ratio (PSNR) JPEG JPEG Vidware Compression ratio (%) (b) Fig. 4. (a )SSIM and (b) PSNR comparison among JPEG, JPEG, and Vidware Vision TM 3. No-Reference Quality Measures For fair comparisons, the computed sharpness values are first normalized with respect to the corresponding values of the uncompressed images. Figure 5 shows the performance comparison among JPEG, JPEG, and Vidware Vision TM based on the proposed measure, NSRT and the Tenengrad measure. In Fig. 5(a), we could see similar output based on full-reference measures. JPEG produces the best performance. For most of the tested compression ratios, the performance of Vidware Vision TM falls in between those of JPEG and JPEG. Eceptions are observed when the compression ratio is larger than approimately 1%, where JPEG outperforms Vidware Vision TM and yields a performance comparable to that of JPEG.

8 758 C.-H. Chen et al. Normalized sharpness measure NSRT JPEG.5 JPEG Vidware Compression ratio (%) (a) Normalized sharpness measure Tenengrad JPEG.8 JPEG Vidware Compression ratio (%) (b) Fig. 5. Comparisons among JPEG, JPEG, and Vidware Vision TM sharpness evaluated by (a) NSRT and (b) Tenengrad based on image The conventional Tenengrad measure shown in Fig. 5(b) is also implemented and serves as comparison reference to validate the use of the newly developed NSRT. The Tenengrad measure is tuned to the local activities of image gradients only and disregards either the sources or the visual effects of these local activities. The artifacts from image compression are also regarded as useful local details, resulting in similar sharpness values for all decompressed images with the same compression ratio. As a consequence, simple Tenengrad measure is insufficient and not applicable. Furthermore, at a compression ratio of.8%, as highlighted by a dashed circle in Fig. 5(b), the Tenengrad measure fails because of the overwhelming blocking artifacts. In comparison, the proposed NSRT is able to distinguish artifacts from desired local details and thus is qualified to evaluate the performances of tested compression methods. In general, the performance of Vidware Vision TM is less sensitive and decreases gradually with respect to the decreased compression ratio. In comparison,

9 Objective Image Quality Evaluation for JPEG, JPEG, and Vidware Vision TM 759 the performance of JPEG deteriorates significantly for compression ratios smaller than % and similar behavior occurs to JPEG for compression ratios smaller than.8%. 3.3 Visual Perception Figure 6 illustrates visual perception of JPEG, JPEG, Vidware Vision TM with.7% compression ratio. It is obvious that JPEG has inferior performance, but it is not easy to judge the performance of JPEG and Vidware Vision TM. (a) (b) (c) (d) Fig. 6. Illustration of visual perception with.7% compression ratio. (a) original 9.5 meters face image; (b) JPEG; (c)jpeg ; (d) Vidware Vision TM. 4 Conclusion In this paper, we evaluated three different compression methods, JPEG, JPEG, and Vidware Version TM by eisting metrics including PSNR, SSIM, and Tenengrad. We also proposed a new image sharpness measure, NSRT, for evaluating the sharpness of decompressed images. According to our eperimental results, we had the following observations: (1) In general, for lower compression ratios (<.7%), Vidware Vision TM outperforms JPEG. () There eists an obvious turning point in compression ratio, beyond which the performances of JPEG and JPEG begin to degrade rapidly. In contrast, the performance of Vidware Vision TM is less sensitive to the compression ratio and the performance decrease is almost uniformly distributed in the tested compression ratio range. (3) Compared to conventional sharpness measures, our proposed image sharpness measure is robust to image artifacts introduced by image compression and produces a reliable evaluation of the sharpness of decompressed images.

10 76 C.-H. Chen et al. Acknowledgments. This work is supported by the University Research Program in Robotics under grant DOE-DE-FG5-4NA5589 and by the DOD/RDECOM/ NAC/ARC Program under grant W56HZV The authors would like to also thank Vidware Incorporated and Databolts Incorporated for kindly providing the compression software used in this study. References 1. JPEG : Verification Model 4., ISO/IEC JTC 1/SC 9/WG 1, Charilaos Christopoulos (Ericsson, Sweden), Editor, April (1999).. JPEG-LS: ISO/IEC : Information Technology-Lossless and near-lossless compression continuous-tone still images. 3. Vidware Vision TM website, 4. Databolts TM website, 5. Zhu, L., Zhang, A., Rao, A., and Srihari, R.: Keyblock: An Approach for Content-Based Image Retrieval. ACM Press New York (). 6. Zunino, R. and Rovetta, S.: Vector Quantization for License-Plate Location and Image Coding, IEEE Trans. on Industrial Electronics, Vol. 47, No. 1, pp , Feb. (). 7. Dang, P. P. and Chau, P. M.: Image Encryption for Secure Internet Multimedia Applications. IEEE Transactions on Consumer Electronics, Vol. 46, No. 3, pp , Aug. (). 8. Lambrecht, C. J. B. and Verscheure, O.: Perceptual Quality Measure Using a Spatial- Temporal Model of Human Visual System. Digital Video Compression Algorithms and Technologies, Proc. SPIE, Vol. 668, San Jose, (1996). 9. Eskicioglu, A. M. and Fisher, P. S.: Image Quality Measures and Their Performance. IEEE Transactions on Communications, Vol. 43, No. 1, Dec. (1995). 1. Girod, B.: What s Wrong with Mean-Squared Error in Digital Images and Human Vision. MA: MIT Press, pp. 7-, (1993). 11. Christopoulos, C., Skodras, A., and Ebrahimi, T.: The JPEG still image coding system: An Overview. IEEE Trans. on Consumer Electronics, Vol. 46, No. 4, pp , Nov. (). 1. Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P.: Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE Trans. on Image Processing, Vol. 13, No. 4, pp. 6-61, April (4). 13. Santos, A., de Solorzano, C. O., Vaquero. J. J., Pena, J. M., Malpica, N., and del Pozo, F.: Evaluation of Autofocus Functions in Molecular Cytogenetic Analysis. Journal of Microscopy, vol. 188, pp. 64-7, Dec. (1997). 14. Ramponi, G.: A Cubic Unsharp Masking Technique for Contrast Enhancement.: Signal Processing, vol. 67, no., pp. 11-, Jun. (1998). 15. Ramponi, G. and Polesel, A.: Rational Unsharp Masking Technique. Journal of Electronic Imaging, vol. 7, no., pp , April (1998). 16. Yao, Y., Abidi. B. R., and Abidi, M. A.: Digital Imaging with Etreme Zoom: System Design and Image Restoration. IEEE Conf. on Computer Vision Systems, New York, Jan. (6).