PERFORMANCE EVALUATION OF 3D IMAGE COMPRESSION

TECHNIQUES
GANAVI C.N
M.Tech (Electronics), Dept. of ECE, HKBK college of Engineering, Bangalore

AIJAZ ALI KHAN

Associate Professor, Dept. of ECE, HKBK college of Engineering, Bangalore

ABSTRACT: Three dimensional(3D) imagesare used in various fields like

telemedicine,telecommunication, satellite communication etc. Compression of these 3D images is
difficult due to strong inter and intraband dependencies that are present along the spatial and
spectral directions. Hence an efficient compression technique is necessary for compression of
these images.There are many new and powerful techniques that are developed for the compression
of 3D images. One such compression technique for the compression of 3D images like medical
images has been developed in ISRO Satellite center. This algorithm uses wavelet based technique
for compression. The performance of the developed algorithm i.e. Once Encoding Selective
Decoding (OESD) with the state of art algorithms like Joint Pictures Experts Group (JPEG2000)
and Three Dimensional Set Partitioning of Hierarchical Tress (3D SPIHT) is presented in this
paper. The OESD algorithm possess features like random access, scalability and Region of
Interest (ROI) which are important for progressive transmission of bit streams which helps for fast
and easy access of bits with high quality of the reconstructed image. The performance is studied
by varying various parameters that are responsible for effective compression of 3D images like
hyperspectral and medical images.

KEYWORDS: 3D images (Three dimensional images), OESD (Once Encoding Selective

Decoding), 3D SPIHT (Set Partitioning of Hierarchical Tress), JPEG2000 (Joint Pictures Experts
Group), hyperspectral images, medical images, progressive transmission, random access, region of
interest, scalability

INTRODUCTION

Images are nothing but an artifact that depicts the visual perception whatever the eyes of human
beings record. Images are two dimensional where the pixels are present in horizontal and vertical
direction. Nowadays 3D (volumetric) images have gained much popularity and play a key role in
many fields like in disaster assessment and management, medical applications, geography,
classification of agriculture etc. The pixels are present along third dimension in addition to height
and width in case of 3D images. Some of the examples of 3D images are remote sensing images
like multispectral images, hyperspectral images, medical images etc.

Large amount of data is being captured and transmitted back to earth by imaging sensors which are
present in satellites. As the data of these images that are captured by sensors are huge, access,
transport and storage is difficult. Therefore there is a need to reduce the size of the data in order to
match the required bandwidth and to store. Hence compression of images is more essential in
order to reduce the redundancy to use space for storage effectively to reduce transmission time and
transmission bandwidth.In this paper focus is given on hyperspectral and medical MR images. 3D
images can also be considered as a sequence of 2D images or slices. Compression of 3D images is
done by directly applying a 2D compression algorithm to each slice separately. The slices are
highly correlated with one another. Therefore a transform must be applied to decorrelate the data.
Hyperspectral images are obtained by collecting hundreds of narrow bands of spectral data which
are generated by imaging sensors. Fast access of particular region is important for many
applications so that low resolution region can be transmitted fast or only some part of the spectrum
can be accessed faster. Also scalability and random access is necessary for many purposes.
Medical images are also three dimensional images where spatial data is present in all the three
dimensions. Medical images are used in fields like telemedicine networks. They require efficient
compression without any loss of information. Several diagnostic imaging techniques such as
magnetic resonance (MR), positron emission tomography (PET) and single photon emission
computed tomography (SPECT) which produces 3D images which are represented as multiple 3D
slices. ROI compression of medical images is necessary in order to retrieve quickly some portion
of the image.A lossless or nearly lossless compression is required in order to maintain the quality
of these images.

Current volumetric image compression schemes mainly exploit spatial and spectral correlations
along different directions. These techniques include vector quantization (VQ), predictive coding,
principle component analysis (PCA), look up table (LUT) etc. All the above techniques provide
efficient compression. However none of them generates embedded bit streams and therefore
cannot provide progressive transmission. To include the embedded bit stream requirement and to
obtain better lossless compression performance, transform coding was introduced. The main aim
of the transform is to jointly exploit the spectral and spatial data correlations.

Therefore in this work, focus is given on evaluating the robustness of the compression techniques
for lossless compression of 3D images. This paper is organized as follows section 2 provides a
brief overview of three different wavelet based compression techniques namely 3D SPIHT,
JPEG2000 and OESD. Section 3 the design of GUI for the comparison of three techniques is
explained. In Section 4 the experimental results are presented. Finally conclusion about the result
is given in section 5.

MATERIALS AND METHOD

Image Selection
In this paper, volumetric images like MR images and hyperspectral images data is considered. The
data consists of a set of 32 sagittal MR images of head obtained from NIMHANS, Bangalore.
Each MRI image dataset is 256 x 256 pixels. In addition, each pixel is represented by 8 bits. Also
hyperspectral images which were obtained from ISROs IMS-1 satellite are considered. Each
hyperspectral image consists of 64 bands with 256 x 256 pixels and each pixel is represented by 16
bits. These images are passed through three different coders like 3D SPIHT, OESD and JPEG2000
encoders and the results are analyzed.

Coding Techniques
Three dimensional set partitioning in hierarchical trees (3D SPIHT)
3D SPIHT is the extension of SPIHT technique which was proposed by Said et al (2008). 3D
SPIHT possess features like random access, resolution scalability and progressive transmission.
The advantages of this technique that the transmission errors will be reduced since only the region
of interest is transmitted without transmitting the whole image. Also the memory required is

586
reduced since the processing is done only on one part of the image, the number of coefficients
involved is reduced so that memory needed to store the lists is also reduced. The encoding is as
follows, the 3D wavelet decomposition is performed by applying sequentially 1D transform across
the three dimensions. The 3D wavelet transform is first applied along spectral direction and then
along the spatial direction. A tree block is formed by 8 coefficients from the lowest subband
forming 2x2x2 cube with all the descendants. The spectral link in the spectral direction is kept
only for the lower spatial subbands [5]. The tree block is encoded according to the SPIHT
algorithm were the wavelet coefficients are scanned column wise then row wise starting from the
lowest subbands to highest subbands. A threshold is selected based on the coefficient which has
largest magnitude. The coding is done by two passes first sorting pass and then refinement pass.
During the sorting pass the coefficients present in list of insignificant pixels (LIP) are sorted and
those coefficients are sent to listof significant pixels (LSP) along with the sign, if they are
significant. The descendants of the significant coefficients are checked and they are moved to LSP
if they are significant or else they are moved to list of insignificant sets (LIS). In refinement pass
the significant bits (i.e. bits of coefficients present in LSP) are sent as output. Then the threshold is
decremented until all the coefficients are encoded. Resolution scalability is done by maintaining
three lists for each resolution. The bit streams are arranged in different order in a block structure so
the any resolution can be obtained at thedecoder. Signal to noise (SNR) scalability is done by
describing a block header each of the size of the bit stream in order to obtain the desired resolution
in all bit planes. The ROI is selected during encoding by adjusting the number of bit planes to be
encoded for a specific block.

Joint Photographic Expert Group (JPEG2000)

JPEG2000 is one of the standards for image compression. JPEG2000 possess features like
progressive transmission of bitstreams, pan/zoom the image, rotating and cropping the image,
ROI, SNR and resolution scalability etc. The advantages are high compression efficiency, ability
to handle large images, error resilience [6]. The drawbacks are the coding algorithm is complex
and computational needs are higher. The main steps involved in JPEG2000 are as follows. First
the color components are divided into tiles (rectangular, non-overlapping regions). Each tile is
compressed separately. Then a wavelet transform is applied on each tile. The wavelet coefficients
are then quantized. The encoding of the wavelet coefficients is done using Embedded Block
Coding with Optimized Truncation (EBCOT) algorithm proposed by Taubman (2000). EBCOT
divides each subband of wavelet coefficients into code blocks and these code blocks are coded
individually. The coding passes in EBCOT are forward and reverse significance propagation pass,
magnitude refinement pass, normalization pass. The bits from code blocks are stored in the form
of packets. The last step is used to construct bitstreams using the packets. ROI can be specified by
using markers for the bitstreams. The bitstreams are organized in the form of layers for resolution
scalability and progressive transmission.

Once Encoding Selective Decoding (OESD)

Unlike 3D SPIHT and JPEG2000, OESD also possess features like spatial, SNR scalability and
ROI based coding. These parameters are defined at the time of decoding. The advantages of OESD
are that the encoded bitstreams that are produced by the four lists like SPIHT can be merged to
single list so that only one pointer is used to select the ROI. This reduces the coding overhead.
Also there is no overlapping of the wavelet coefficients since the image is segmented only along
the spatial direction.The steps in encoding are as follows. For 3D image, a 2D mask is generated
by using histogram then the background is marked zero (Do not care). The mask is divided into 3
segments. 3D integer lifting wavelet transform is applied to each segment. These masks are

587
multiplied with the original image.The 3D image is also titled like JPEG2000 into 3 segments.
Then 3D integer lifting wavelet transform is applied to each segment first in spatial direction then
along rows and columns.The 3D image along with the mask is divided into 3D lattice of size r x c
x t, where r and c are the rows and columns; t is the spatial or temporal direction.For each of the
3D lattice k, Wmax is determined based on the maximum of the wavelet coefficients. Threshold T
of k is computed as T = (log10 (Wmax)). The maximum of T is stored in T(g).The wavelet
coefficients are scanned from lowest L band .Within each subband the coefficients are first
scanned in zigzag manner like SPIHT according to their significance. The 3D lattices are scanned
in raster scan manner like JPEG2000 first along the rows and columns then along the temporal
direction.Significant pass: Check for the significance of the 3D lattice; it is significant, if at least
one of its coefficients is greater than or equal to 2T (g). If it is not significant, store 0 in list lis
(indicating lattice insignificance). If the 3D lattice is first time significant, store 1 in the list lis.
Else, if the lattice is already significant, do not store anything.For each of the significant
coefficient in each subband; check for the descendants, if the coefficient is significant, store 1 in
the list lis. Also, if the coefficient is positive significant, store 0 in lis; otherwise store.
Refinement pass: Store the current bit of all the significant coefficients (at the previous threshold)
in the lis. If T (g) is 0, go to next step; otherwise, set T (g) =T (g) 1 and then repeat the
scanning. After completing all the bands of the current wavelet resolution, store the next index of
the list 'lis' in index_list. Repeat steps for the next resolution. The output obtained is arithmetic
coded. Along with the list lis, the threshold array, image size, and the chain coded image mask
are required to be sent to the decoder.Decoding can be performed by simply reversing the
operations corresponding to encoding.

EXPERIMENTAL RESULTS

To compare the 3D SPIHT, OESD and JPEG2000 (part 1), a Graphical User Interface (GUI) is
built in MATLAB 2012a. The GUI consists of the selection box; compare summary table and
message box. There is a provision for the user to change the type of wavelet (except for JPEG2000
where the wavelet is 5/3 for lossless), the number of decomposition levels. The compare summary
table is used for the display of the results. Message box displays the error that occurs during
execution. The comparison is made by specifying the different wavelets and decomposition levels
for each of the algorithm for medical images as well as hyperspectral images.

We use compression ratio (Cr) and bits per pixel (Bpp) to evaluate the performances of the three
coders. Cr is the ratio of number of bits required to represent the data before compression to the
number of bits required to represent the data after compression. Generally higher the compression
ratio poorer is the quality of the resulting image. Cr is given by

Compression ratio (Cr) = (1)

Bpp is the number of bits needed on average to compress one pixel of image. Bpp is given by
Bits Per Pixel (Bpp) = (2)

588
 Figure 2: GUI layout in MATLAB

Table 1: Results when nimhans filtered images were used for three different techniques for
decomposition levels 1 and 2 spatially and spectral level 1
3D SPIHT OESD JPEG2000
Wavelets Decomposition Levels
Cr Bpp Cr Bpp Cr Bpp
(2,2) 1.9010 4.2080 3.5313 2.3180
(5,3) 1 1.7783 4.4987 3.2304 2.4765 2.6470 3.0222
Haar 1.7585 4.5494 3.5392 2.2604
(2,2) 3.9428 2.0291 3.4324 2.3307
(5,3) 2 3.8314 2.0880 3.1383 2.5492 2.5955 3.0822
Haar 3.6941 2.1656 3.6211 2.2092

Table 2: Results when hyperspectral images were used for three different techniques for
decomposition levels 1 and 2 spatially and spectral level 1
3D SPIHT OESD JPEG2000
Wavelets Decomposition Levels
Cr Bpp Cr Bpp Cr Bpp
(2,2) 1.6606 4.8175 3.0123 2.6478
(5,3) 1.9505 4.1015 2.9644 2.6987 10.825 0.73901
1
Haar 1.5612 5.1243 3.1100 2.5724
(2,2) 3.5985 2.2231 3.2315 2.4756
(5,3) 3.5313 2.2655 3.1098 2.5725 14.000 0.5714
2
Haar 3.3507 2.3876 3.3654 2.3771

589
CONCLUSION

The results show that the OESD performs better than JPEG2000 in case of MR images and the
results of OESD are comparable with 3D SPIHT in case of hyperspectral images for
decomposition levels 1 and 2. The compression ratio that is obtained for MR images from OESD
is more than 3D SPIHT for different types of wavelets. It is observed from the tables that Haar
performs well in case of OESD. The compression ratio increases from level 1 to level 2 in case of
3D SPIHT but as the compression ratio increases the quality of the reconstructed image is poor. In
case of OESD and JPEG2000 the compression ratio decreases by 0.1% to 0.5% and the bits per
pixel increases by 0.1% to 0.5% bpp. Therefore from results it can be observed that OESD is
stable technique for both hyperspectral and medical MR images.

REFERENCES

V. Ramesh Naidu, V.K Agarwal, A.G Ramakrishnan, Once Encoding and Selective Decoding of
3D MRI with SNR and Spatial Scalability, Managing complexity in a distributed world, 27th
May-31 May 2008, IISC, Bangalore
Ying Hou, Lossy to lossless compression of hyperspectral image using the 3D set partitioning
embedded zero Block coding algorithms, Journal on software engineering & applications, July
2009.
Gloria Menegazand Jean-Philippe Thiran, Three dimensional encoding/two dimensional
decoding of medical data, IEEE transactions on medical imaging, volume 22, no. 3, March
2003.
R. Srikanth, A.G. Ramakrishnan, Region of Interest Based Coding of 2-D and 3-D Magnetic
Resonance Images, Indian Institute of Science, Bangalore.
E. Christophe and W. A. Pearlman, Three-dimensional SPIHT coding of volume images with
random access and resolution scalability, EURASIP J. Image Video Process, 2008.
D. Taubman, High performance scalable image compression with EBCOT, IEEE Trans. Image
Process., vol. 9, no. 7, pp. 1158-1170, Jul. 2000.

590