Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition, 1st Edition (2015)
Part I. Image and Signal Processing
Chapter 11. ρGBbBShift
Method for introducing perceptual criteria to region of interest coding
Jaime Moreno1,2; Oswaldo Morales1; Ricardo Tejeida1 1 Superior School of Mechanical and Electrical Engineers, National Polytechnic Institute of Mexico, IPN Avenue, Lindavista, Mexico City, Mexico
2 Signal, Image and Communications Department, University of Poitiers, Poitiers, France
Abstract
This work describes perceptual generalized bitplane-by-bitplane shift (ρGBbBShift) a perceptual method for coding of region of interest (ROI) and it is based on Moreno et al. (2013) [1]. Then, this article introduces perceptual criteria to the GBbBShift method when bit planes of ROI and background areas are shifted. This additional feature is intended for balancing perceptual importance of some coefficients regardless of their numerical importance. Perceptual criteria are applied using a contrast band-pass filtering, which is a low-level computational model that reproduces color perception in the human visual system. Results show that there is no perceptual difference at ROI between the MaxShift method and ρGBbBShift and, at the same time, perceptual quality of the entire image is improved when using ρGBbBShift. Furthermore, when ρGBbBShift method is applied to Hi-SET coder and it is compared against MaxShift method applied to both the JPEG2000 standard and the Hi-SET, the images coded by the combination ρGBbBShift-Hi-SET get the best results when the overall perceptual image quality is estimated. The ρpGBbBShift method is a generalized algorithm that can be applied to other Wavelet-based image compression algorithms such as JPEG2000, SPIHT, or SPECK.
Keywords
Bitplane coding
Bitplane-by-bitplane shift
Generalized bitplane-by-bitplane shift
Hi-SET
Image coding
JPEG2000
Maximum shift
Region of interest
Wavelet coding
ACKNOWLEDGMENT
This work is supported by National Polytechnic Institute of Mexico by means of Project No. 20140096, the Academic Secretary and the Committee of Operation and Promotion of Academic Activities (COFAA), National Council of Science and Technology of Mexico by means of Project No. 204151/2013, and LABEX Σ-LIM France, Coimbra Group Scholarship Programme granted by University of Poitiers and Region of Poitou-Charentes, France.
1 Introduction
Region of interest (ROI) image coding is a feature that modern image coders have, which allows to encode n specific region with better quality than the rest of the image or background (BG). ROI coding is one of the requirements in the JPEG2000 image coding standard [2,3], which defines two ROI methods [4,5]:
(1) Based on general scaling [4]
(2) Maximum shift (MaxShift) [5]
The general ROI scaling-based method scales coefficients in such a way that the bits associated with the ROI are shifted to higher bitplanes than the bitplanes associated with the BG, as shown in Figure 1(b). It implies that during an embedded coding process, any BG bitplane of the image is located after the most significant ROI bitplanes into the bitstream. But, in some cases, depending on the scaling value, φ, some bits of ROI are simultaneously encoded with BG. Therefore, this method allows to decode and refine the ROI before the rest of the image. No matter φ, it is possible to reconstruct with the entire bitstream a highest fidelity version of the whole image. Nevertheless, if the bitstream is terminated abruptly, the ROI will have a higher fidelity than BG.
FIGURE 1 JPEG2000ROI Coding. (a) NoROI coding, (b) scaling-based ROI coding method (φ = 3), and (c) MaxShift method, φ = 7. Background is denoted as BG, region of interest as ROI, and bitplane mask as BPmask. MSB is the most significant bitplane and LSB is the least significant bitplane.
The scaling-based method is implemented in five steps:
(1) A wavelet transform of the original images is performed.
(2) An ROI mask is defined, indicating the set of coefficients that are necessary for reaching a lossless ROI reconstruction, Figure 2.
FIGURE 2 ROI mask generation, wavelet domain.
(3) Wavelet coefficients are quantized and stored in a sign magnitude representation, using the most significant part of the precision. It will allow to downscale BG coefficients.
(4) A specified scaling value, 
, downscales the coefficients inside the BG.
(5) The most significant bitplanes are progressively entropy encoded.
The input of ROI scaling-based method is the scaling value φ, while MaxShift method calculates it. Hence, the encoder defines from quantized coefficients this scaling value such that,
(1)
where max{MBG} is the maximum coefficient in the BG. Thus, when ROI is scaled up φ bitplanes, the minimum coefficient belonging to ROI will be place one bitplane up of BG (Figure 1(c)). Namely, 2φ is the smallest integer that is greater than any coefficient in the BG. MaxShift method is shown in Figure 1(c). Bitplane mask (BPmask) will be explained in Section 2.1.
At the decoder side, the ROI and BG coefficients are simply identified by checking the coefficient magnitudes. All coefficients that are higher or equal than the φth bitplane belong to the ROI, otherwise, they are a part of BG. Hence, it is not important to transmit the shape information of the ROI or ROIs to the decoder. The ROI coefficients are scaled down φ bitplanes before inverse wavelet transformation is applied.
2 Related work
2.1 BbB Shift
Wang and Bovik proposed the bitplane-by-bitplane shift (BbBShift) method in Ref. [6]. BbBShift shifts bitplanes on a bitplane-by-bitplane strategy. Figure 3(a) shows an illustration of the BbBShift method. BbBShift uses two parameters, φ1 and φ2, whose sum is equal to the number of bitplanes for representing any coefficient inside the image, indexing the top bitplane as bitplane 1. Summarizing, the BbBShift method encodes the first φ1 bitplanes with ROI coefficients, then, BG and ROI bitplanes are alternately shifted, refining gradually both ROI and BG of the image (Figure 3(a)). The encoding process of the BbBShift method is defined as
FIGURE 3 ROI coding methods. (a) BbBShift, φ1 = 3 and φ2 = 4, (b) GBbBShift and (c) ρGBbBShift. Background is denoted as BG (for ρGBbBShift method is perceptually quantized by ρSQ at d2), region of interest as ROI (for ρGBbBShift method is perceptually quantized at d1 by ρSQ) and bitplane mask as BPmask.
(1) For a given bitplane bpl with at least one ROI coefficient:
• If bpl ≤ φ1, bpl is not shifted.
• If φ1 < bpl ≤ φ1 + φ2, bpl is shifted down to φ1 + 2 (bpl − φ1)
(2) For a given bitplane bpl with at least one BG coefficient:
• If bpl ≤ φ2, bpl is shifted down to φ1 + 2bpl − 1
• If bpl > φ2, bpl is shifted down to φ1 + φ2 + bpl
2.2 GBbBShift
In practice, the quality refinement pattern of the ROI and BG used by BbBShift method is similar to the general scaling-based method. Thus, when the image is encoded and this process is truncated in a specific point, the quality of the ROI is high while there is no information of BG.
Hence, Wang and Bovik [7] modified BbBShift method and proposed the generalized bitplane-by-bitplane shift (GBbBShift) method, which introduces the option to improve visual quality either of ROI or BG or both. Figure 3(b) shows that with GBbBShift method it is possible to decode some bitplanes of BG after the decoding of same ROI bitplanes. It allows to improve the overall quality of the recovered image. This is possible gathering BG bitplanes. Thus, when the encoding process achieves the lowest bitplanes of ROI, the quality of BG could be good enough in order to portray an approximation of BG.
Therefore, the main feature of GBbBShift is to give the opportunity to arbitrary chose the order of bitplane decoding, grouping them in ROI bitplanes and BG bitplanes. This is possible using a binary bitplane mask or BPmask, which contains one bit per each bitplane, that is, twice the amount of bitplanes of the original image. An ROI bitplane is represented by 1, while a BG bitplane by 0. For example, the BPmask for MaxShift method in Figure 1(c) is 11111110000000, while for BbBShift in Figure 3(a) and GBbBShift inFigure 3(b) are 11101010101000 and 11100011110000, respectively.
At the encoder side, the BPmask has the order of shifting both the ROI and BG bitplanes. Furthermore, BPmask is encoded in the bitstream, while the scaling values φ or φ1 and φ2 from the MaxShift and BbBShift methods, respectively, have to be transmitted.
3 Perceptual GBbBShift
3.1 Quantization
In order to generate an approximation to how every pixel is perceived from a certain distance taking into account the value of its neighboring pixels the Chromatic Induction Wavelet Model (CBPF) is used. CBPF attenuates the details that the human visual system (HVS) is not able to perceive, enhances those that are perceptually relevant and produces an approximation of the image that the brain visual cortex perceives. CBPF takes an input image I and decomposes it into a set of wavelet planes ωs,o of different spatial scales s (i.e., spatial frequency ν) and spatial orientations o. It is described as:
(2)
where n is the number of wavelet planes, cn is the residual plane, and o is the spatial orientation either vertical, horizontal, or diagonal. The perceptual image Iρ is recovered by weighting these ωs,o wavelet coefficients using the extended Contrast Sensitivity Function (e-CSF), which considers spatial surround information (denoted by r), visual frequency (ν related to spatial frequency by observation distance), and observation distance (d). Perceptual image Iρ can be obtained by
(3)
where α(ν,r) is the e-CSF weighting function that tries to reproduce some perceptual properties of the HVS. The term α(ν,r) ωs,o ≡ ωs,o;ρ,d can be considered the perceptual wavelet coefficients of image I when observed at distance d. For details on the CBPF and the α(ν,r) function, see Ref. [8].
We employ the perceptual quantizer (ρSQ) either forward (F-ρSQ) and inverse (I-ρSQ), defined by Moreno et al. [1]. Each transform sample at the perceptual image Iρ (from Equation 3) is mapped independently to a corresponding step size either Δs or Δn, thus Iρis associated with a specific interval on the real line. Then, the perceptually quantized coefficients Q(F-ρSQ), from a known viewing distance d, are calculated as follows:
(4)
The perceptual inverse quantizer (I-ρSQ) or the recovered 
introduces perceptual criteria to the classical Inverse Scalar Quantizer and is given by
(5)
3.2 ρGBbBShift Algorithm
In order to have several kinds of options for bitplane scaling techniques, a perceptual generalized bitplane-by-bitplane shift (ρGBbBShift) method is proposed. The ρGBbBShift method introduces to the GBbBShift method perceptual criteria when bitplanes of ROI and BG areas are shifted. This additional feature is intended for balancing perceptual importance of some coefficients regardless of their numerical importance and for not observing visual difference at ROI regarding MaxShift method, improving perceptual quality of the entire image.
Thus, ρGBbBShift uses a binary bitplane mask or BPmask in the same way that GBbBShift (Figure 3(c)). At the encoder, shifting scheme is as follows:
(1) Calculate φ using Equation (1).
(2) Verify that the length of BPmask is equal to 2φ.
(3)
• For all ROI Coefficients, forward perceptual quantize them using Equation (4) (F-ρSQ) with viewing distance d1.
• For all BG coefficients, forward perceptual quantize them using Equation (4) (F-ρSQ) with viewing distance d2, being d2 
d1.
(4) Let τ and η be equal to 0.
(5) For every element i of BPmask, starting with the least significant bit:
• If BPmask(i) = 1, Shift up all ROI perceptual quantized coefficients of the (φ − η)th bitplane by τ bitplanes and increment η.
• Else: Shift up all BG perceptual quantized coefficients of the (φ − τ)th bitplane-by-η-bitplanes and increment τ.
At the decoder, shifting scheme is as follows:
(1) Let φ = length of BPmask/2 be calculated.
(2) Let τ and η be equal to 0.
(3) For every element i of BPmask, starting with the least significant bit:
• If BPmask(i) = 1, Shift down all perceptual quantized coefficients by τ bitplanes, which pertain to the (2φ − (τ + η))th bitplane of the recovered image and increment η.
• Else: Shift down all perceptual quantized coefficients by η bitplanes, which pertain to the (2φ − (τ + η))th bitplane of the recovered image and increment τ.
(4) Let us denote as ci,j a given non-zero wavelet coefficient of the recovered image with 2φ bitplanes and ci,j as a shifted down c obtained in the previous step, with φ bitplanes:
• If (ci,j& BPmask) > 0, inverse perceptual quantize 
using Equation (5) (I-ρSQ) with d1 as viewing distance.
• If (ci,j& BPmask) = 0, inverse perceptual quantize using Equation (5) (I-ρSQ) with d2 as viewing distance.
4 Experimental results
The ρGBbBShift method, as the other methods presented here, can be applied to many image compression algorithms such as JPEG2000 or Hi-SET [9]. We test our method applying it to Hi-SET and the results are contrasted with MaxShift method in JPEG2000 and Hi-SET. The setup parameters are φ = 8 for MaxShift and BPmask = 1111000110110000, d1 = 5H and d2 = 50H, where His picture height (512 pixels) in a 19-in. LCD monitor, for ρGBbBShift. Also, we use the JJ2000 implementation when an image is compressed by JPEG2000 standard [10].
4.1 Application in Well-Known Test Images
Figure 4 shows a comparison among methods MaxShift and GBbBShift applied to JPEG2000, in addition to, ρGBbBShift applied to Hi-SET. The 24-bpp image Barbara is compressed at 0.5 bpp. It can be observed that without visual difference at ROI, the ρGBbBShift method provide better image quality at the BG than the general based methods defined in JPEG2000Part II [2].
FIGURE 4 512 × 640 pixel Image Barbara with 24 bpp. ROI is a patch of the image located at [341 280 442 442], whose size is 1/16 of the image. Decoded images at 0.5 bpp using MaxShift method in JPEG2000 coder ((a) φ = 8), GBbBShift method in JPEG2000 coder ((b) BPmask = 1111000110110000), and ρGBbBShift method in Hi-SET coder ((c) BPmask = 1111000110110000). (a) MaxShift in JPEG2000 coder, 0.5 bpp. (b) GBbBShift in JPEG2000 coder, 0.5 bpp. (c) ρGBbBShift in Hi-SET coder, 0.5 bpp.
In order to better qualify the performance of MaxShift, GBbBShift, and ρGBbBShift methods, first, we compared these methods applied to the Hi-SET coder and then, we compare MaxShift and ρGBbBShift methods applied to the JPEG2000 standard and Hi-SET, respectively. We compress two different grayscale and color images of 1600, from CSIQ image database [11], and Lenna [12] at different bit-rates. ROI area is a patch at the center of these images, whose size is 1/16 of the image. We employ the perceptual quality assessment proposed by Moreno et al. [13] called P2SNR, which weights the mainstream PSNR by means of a chromatic induction model, so we renamed this image quality assessment as CwPSNR.
Figure 5 shows the comparison among MaxShift (Blue Function), GBbBShift (Green Function), and ρGBbBShift (Red Function) methods applied to Hi-SET coder. 512 × 512 pixel Image 1600 both for grayscale and color are employed for this experiment. These figures also show that the ρGBbBShift method gets the better results both in PSNR (objective image quality) and CwPSNR (subjective image quality) in contrast to MaxShift and GBbBShift methods.
FIGURE 5 Comparison among MaxShift (Blue Function), GBbBShift (Green Function) and ρGBbBShift (Red Function) methods applied to Hi-SET coder. 512 × 512 pixel Image 1600 with (a-b) 8 and (c-d) 24 bpp are employed for this experiment. ROI is a patch at the center of the image, whose size is 1/16 of the image. The overall image quality of decoded images at different bits per pixel is contrasted both (a and c) objectively and (b and d) subjectively. (a) PSNR grayscale. (b) CwPSNR grayscale. (c) PSNR color. (d) CwPSNR color.
When MaxShift method applied to JPEG2000 coder and ρGBbBShift applied to Hi-SET coder are compared, in the whole image quality assessment of image 1600, JPEG2000 obtains better objective quality both for grayscale and color images (Figure 6(a) and (c), respectively). But when the subjective quality is estimated ρGBbBShift coded images are perceptually better.
FIGURE 6 Comparison between MaxShift method applied to JPEG2000 coder and rGBbBShift applied to Hi-SET coder. 512 x 512 pixel Image 1600 with (a-b) 8 and (c-d) 24 bits per pixel are employed for this experiment. ROI is a patch at the center of the image, whose size is 1/16 of the image. The overall image quality of decoded images at different bits per pixel are contrasted both (a and c) objectively and (b and d) subjectively.
A visual example is depicted by Figure 7, where it can be shown that there is no perceptual difference between ROI areas besides the perceptual image quality at BG is better when ρGBbBShift is applied to the Hi-SET coder (Figure 7(d)). Furthermore, Figure 7(b) and (c) shows the examples when MaxShift and GBbBShift methods, respectively, are applied to the Hi-SET coder.
FIGURE 7 512 × 512 pixel Image 1600 from CSIQ image data base with 8 bpp. ROI is a patch at the center of the image, whose size is 1/16 of the image. Decoded images at 0.42 bpp using φ = 8 for MaxShift method (a) in JPEG2000 coder and (b) in Hi-SET coder, and BPmask = 1111000110110000 for (c) GBbBShift and (d) ρGBbBShift methods in Hi-SET coder.
Similarly, when an ROI area is defined in Image Lenna, Figure 8(a) and (b) shows the comparison among MaxShift (Blue Function), GBbBShift (Green Function), and GBbBShift (Red Function) methods applied to Hi-SET coder. 512 × 512 pixel Image Lenna for grayscale is employ for this experiment. These figures also show that the ρGBbBShift method gets the better results both in PSNR (objective image quality, Figure 8(a)) and CwPSNR (subjective image quality, Figure 8(b)) in contrast to MaxShift and GBbBShift methods. In addition, when MaxShift method applied to JPEG2000 coder and ρGBbBShift applied to Hi-SET coder are compared, ρGBbBShift obtains less objective quality (Figure 8(c)), but better subjective quality for grayscale images (Figure 8(d)).
FIGURE 8 (a and b) Comparison among MaxShift (Blue Function), GBbBShift (Green Function) and ρGBbBShift (Red Function) methods applied to Hi-SET coder. (c and d) Comparison between MaxShift method applied to JPEG2000 coder and ρGBbBShift applied to Hi-SET coder. 512 × 512 pixel Image Lenna with 8 bpp is employed for this experiment. ROI is a patch at the center of the image, whose size is 1/16 of the image. The overall image quality of decoded images at different bits per pixels are contrasted both (a and c) objectively and (b and d) subjectively.
Figure 9 shows a visual example, when image Lenna is compressed at 0.34 bpp by JPEG2000 and Hi-SET. Thus, it can be observed that ρGBbBShift provides an important perceptual difference regarding the MaxShift method (Figure 9(d)). Furthermore, Figure 9(b) and (c) shows the examples when MaxShift and GBbBShift methods, respectively, are applied to the Hi-SET coder.
FIGURE 9 512 512 pixel Image Lenna from CMU image database with 8 bpp. ROI is a patch at the center of the image, whose size is 1/16 of the image. Decoded images at 0.34 bpp using φ = 8 for MaxShift method (a) in JPEG2000 coder and (b) in Hi-SET coder, and BPmask = 1111000110110000 for (c) GBbBShift and (d) ρGBbBShift methods in Hi-SET coder.(a) MaxShift method in JPEG2000 coder, 0.34 bpp. (b) MaxShift method in Hi-SET coder, 0.34 bpp. (c) GBbBShift method in Hi-SET coder, 0.34 bpp. (d) ρGBbBShift method in Hi-SET coder, 0.34 bpp.
4.2 Application in Other Image Compression Fields
The usage of ROI coded images depends on an specific application, but in some fields such as manipulation and transmission of images is important to enhance the image quality of some areas and to reduce it in others [14,15]. In telemedicine or in remote sensing (RS) it is desirable to maintain the best quality of the ROI area, preserving relevant information of BG, namely, the most perceptual frequencies.
Thus, in medical applications an image is by itself an ROIϕ area of the human body, a mammography is an area of chest, for instance. That is why, it is important to know where is this ROIϕ located, in order to ease the interpretation of a given ROI coded image. In addition, according Federal laws in some countries, ROI areas must be lossless areas [16]. ρGBbBShift is able to accomplish these two features.
Figure 10 shows an example of medical application. A rectangular ROI of the Image mdb202 from PEIPA image database [17], coordinates [120 440 376 696], is coded at 0.12 bpp by JPEG2000 and Hi-SET, employing MaxShift and ρGBbBShift methods, respectively. The overall image quality measured by PSNR in Figure 10(a) (MaxShift method applied to JPEG2000) is 37.21 dB, while in Figure 10(c) (ρGBbBShift method applied to Hi-SET) is 36.76 dB. Again, PSNR does not reflect perceptual differences between images (Figure 10(b) and (d)). When perceptual metrics assess the image quality of the ρGBbBShift coded image, for example, VIFP = 0.6359, WSNR = 34.24 and CwPSNR = 40.88, while for MaxShift coded image VIFP = 0.3561, WSNR = 31.34, and CwPSNR = 37.18. Thus, these metrics predicts that there is an important perceptual difference between ROI methods, being ρGBbBShift method better than MaxShift method.
FIGURE 10 Example of a medical application. 1024 x 1024 pixel Image mdb202 from PEIPA image database. ROI is a patch with coordinates [120 440 376 696], whose size is 1/16 of the image. Decoded images at 0.12 bpp using MaxShift method ((a-b) φ = 8) in JPEG2000 coder and ρGBbBShift method ((c-d) BPmask = 1111000110110000) in Hi-SET coder. (a) MaxShift method in JPEG2000 coder, 0.12 bpp. (b) Patch of (a) portrayed both ROI and BG areas.(c) ρGBbBShift method in Hi-SET coder, 0.12 bpp. (d) Patch of (c) portrayed both ROI and BG areas.
5 Conclusions
A perceptual implementation of the ROI, ρGBbBShift, is proposed, which is a generalized method that can be applied to any wavelet-based compressor. We introduced ρGBbBShift method to the Hi-SET coder and it visually improves the results obtained by previous methods like MaxShift and GBbBShift. Our experiments show that ρGBbBShift into Hi-SET provides an important perceptual difference regarding the MaxShift method into JPEG2000, when it is applied to conventional images like Lenna or Barbara.
References
[1] Moreno J, Fernandez C, Saucedo S. pGBbBShift: method for introducing perceptual criteria to region of interest coding. In: International conference on image processing, computer vision, and pattern recognition of the world congress in computer science, computer engineering, and applied computing; 2013:111–117.
[2] Boliek M, Majani E, Houchin JS, Kasner J, Carlander M, Information technology: JPEG2000 image coding system (extensions), JPEG 2000 part II final committee draft ed., ISO/IEC JTC 1/SC 29/WG 1; December 2000.
[3] Boliek M, Christopoulos C, Majani E, Information technology: JPEG2000 image coding system, JPEG 2000 Part I final committee draft version 1.0 ed., ISO/IEC JTC1/SC29 WG1, JPEG 2000; April 2000.
[4] Taubman DS, Marcellin MW. JPEG2000: image compression fundamentals, standards and practice. Boston, Dordrecht, London: Kluwer Academic Publishers; 2002.079237519X.
[5] Atsumi E, Farvardin N. Lossy/lossless region-of-interest image coding based on set partitioning in hierarchical trees. 87–91. International conference on image processing. 1998;vol. 1.
[6] Wang Z, Bovik AC. Bitplane-by-bitplane shift(Bb BShift)—a suggestion for JPEG2000 region of interest image coding. IEEE Signal Process Lett. 2002;9(5):160–162.
[7] Wang Z, Banerjee S, Evans BL, Bovik AC. Generalized bitplane-by-bitplane shift method for JPEG2000 ROI coding”. 81–84. IEEE international conference on image processing. 2002;vol. 3.
[8] Otazu X, Párraga C, Vanrell M. Toward a unified chromatic induction model. J Vis. 2010;10(12):6.
[9] Moreno J, Otazu X. Image coder based on Hilbert scanning of embedded quad trees: an introduction of Hi-SET coder. In: IEEE international conference on multimedia and expo; 2011.
[10] C. Research, École Polytechnique Fédérale de Lausanne, and Ericsson. JJ2000 implementation in Java. Cannon Research, École Polytechnique Fédérale de Lausanne and Ericsson. [Online]. Available: http://jj2000.epfl.ch/; 2001.
[11] Larson EC, Chandler DM. Most apparent distortion: a dual strategy for full-reference image quality assessment”. Proc SPIE. 2009;742:1–17.
[12] S. I.P. I. of the University of Southern California. The USC-SIPI image database. Signal and Image Processing Institute of the University of Southern California. [Online]. Available: http://sipi.usc.edu/database/; 1997.
[13] Moreno J. P2SNR: perceptual full-reference image quality assessment for JPEG2000. In: International conference on image processing, computer vision, and pattern recognition of the world congress in computer science, computer engineering, and applied, computing; 2012:835–841.
[14] Bartrina-Rapesta J, Auli-Llinas F, Serra-Sagrista J, Zabala-Torres A, Pons-Fernandez X, Maso-Pau J. Region of interest coding applied to map overlapping in geographic information systems. In: IEEE international geoscience and remote sensing symposium; 2007:5001–5004.
[15] Gonzalez-Conejero J, Serra-Sagrista J, Rubies-Feijoo C, Donoso-Bach L. Encoding of images containing no-data regions within JPEG2000 framework. In: 15th IEEE international conference on image processing; 2008:1057–1060.
[16] Wilson B. Ethics and Basic Law for Medical Imaging Professionals. Philadelphia: F.A. Davis Co; 1997.
[17] PEIPA. Pilot European image processing archive, available at http://peipa.essex.ac.uk/ [Online]; 2004.