ITCH - ELECTRO 2007

Octubre 17-19, Chihuahua, México

AN APPROACH TO LOSSY IMAGE COMPRESSION USING 1-D WAVELET TRANSFORMS.

Humberto Lobato-Morales, Otto J. Paz-Luna, Vicente Alarcón-Aquino. Universidad de las Américas-Puebla. Communications and Signal Processing Research Group. Sta. Catarina Mártir, Cholula, Puebla. 72820. México. E-mail: {humberto.lobatoms, ottoj.pazla, vicente.alarcon}@udlap.mx

Summary

The rest of the paper is organized as follows. Section 2 presents a review of wavelet theory; Section 3 describes the proposed algorithm for image compression. Simulation results are reported in Section 4. Conclusions are presented in Section 5.

In this paper, an approach to lossy image compression using 1-D wavelet transforms is proposed. The analyzed image is divided in little subimages and each one is decomposed in vectors following a fractal Hilbert curve. A Wavelet Transform is thus applied to each vector and high frequency components are suppressed. The Huffman coding algorithm is then applied in order to reduce image weight. The 64-point vector transforms used in this work allow lower computational time rather than conventional 8x8-matrix transforms. Two different percentages of high frequency coefficients in one and two level wavelet transform are suppressed achieving good compression ratios and SNR. Exploiting the properties of human visual system, simulation results show that image distortion is less appreciated in image sizes 512 by 512 pixels and greater. Keywords – Discrete Wavelet Transform (DWT), Hilbert curve, coefficient suppression, Huffman coding.

2. Wavelet transform The Wavelet Transform (WT) is a mathematical tool that provides building blocks with information in scale and time of a signal [3]. These building blocks are generated from a single fixed function called mother wavelet by translation and dilation operations. The most commonly used mother wavelets are Haar, Daubechies, Mexican Hat, Morlet, and Walsh [3]. The process of wavelet transform of a signal is called analysis, and the inverse process to reconstruct the analyzed signal is called synthesis. The analysis generates different sub-band blocks (multi-resolution analysis MRA [3]), so different levels can be generated as the application requires. This process is also known as sub-band coding [2]. A discrete one dimensional signal and its three levels of decomposition [2] are shown in figure 1.

1. Introduction A very important task of image processing is image compression and manipulation with low distortions. Standardized image compression formats as GIF and JPEG use sub-images decomposition, bidimensional transforms of each one, and coding algorithms [1] [2]. High frequency components are eliminated at different ratios keeping only the greater energy components, and then a codification method is applied. This coefficient suppression generates a lossy image. Many transforms have been proved for this purpose. Some of them include Discrete Fourier Transform, Discrete Cosine Transform, WalshHadamard Transform, Haar Transform, Hartley Transform, and Wavelet Transform, all of these generating different image distortion and compression values [2]. Recently, several approaches have been proposed for image compression based on wavelet decomposition (see e.g. [6], [7], [8]). By using this sub-image coding and spectral transforms, good compression ratios are obtained using wavelets, achieving low image distortion based on imperfections of human visual systems [1].

Figure 1. SubSub-band coding, a) Original signal and each analysis level level,, b) 3-level DWT.

1

ITCH - ELECTRO 2007

Octubre 17-19, Chihuahua, México

matrix multiplication of the Haar transform matrix Hl by each transposed row of TE:

The resulting signal consists of four sub-bands. The first one corresponds to the higher frequency contents or image details. Sub-band 4 contains the greatest energy components, which correspond to low frequencies. Sub-bands 2 and 3 represent intermediate frequencies. A 2-D DWT is obtained from the 1-D transform of each row of a matrix and then the same procedure is applied to each column. The resulting matrix contains the information in different sub-bands depending on the analysis level.

TFi = H l * TE Ti , i = 1,2,..., m × n / 64 l = 1,2,3.

where l corresponds to the sub-band coding level. The superscript T indicates transpose. The Haar transform matrixes Hl are defined by H1=W1, H2=W2*W1, and H3=W3*W2*W1

3. Proposed approach to image compression Due to special characteristics on the human visual system, it is possible to distort the image without significant degradation of its quality from the observer’s point of view [1]. Based on that fact, high frequency coefficients can be suppressed achieving good compression performances.

1 1 0 0  0 0  0 0 = W1  1 − 1  0 0 0 0  0 0

The first step of proposed compression process is image decomposition in 8x8 sub-images. Each one is then converted to a 64 point-vector following a scan with the Hilbert fractal curve, as it is shown in Figure 2. The main property of using this fractal curve is to follow a more distributed pattern in order to allow more image definition when coefficients are suppressed. The resultant vectors are located in a new matrix called TE with size pxq, where

       W2 =        

(1)

and q=64. The values m and n in (1) compose the original image dimensions. In order to generate a complete TE matrix, images must be adjusted to a square matrix with 2k points by side, where k is a positive integer. A 1-D Discrete Wavelet Transform is realized to each row of the TE matrix, and a new matrix called TF (2) is obtained. The Wavelet used in this work is based on Haar functions, so it is called the Haar Wavelet. The process of sub-band coding is done by averaging and differentiating, which correspond to a low-pass and high-pass filtering along the signal. The normalized filters are ϕ L = 1 2 , 1 2 and ϕ H = 1 2 ,− 1 2 respectively [2]. This analysis technique is realized by

{

}

{

(3)

where Wn, n=1,2,3, contents the basis Haar filters (lowpass and high-pass) for each of the three different frequency bands and positions [2]. W3*W2*W1 contents all the Haar functions for an 8x8 matrix. The values for Wn are:

Figure 2. Standard scan subsub-image.

p = m × n / 64 ,

(2)

      W3 =        

}

2

1

0 1 0

0 0 1

0 0 1

0 0 0

0 0 0 0 0 0 1 −1 0

0 0 0

1 0 0

0 0

1

2

2

0

0

1 2



0 1 0

1 −1 0 0 0 1

0 0 0

0

1

1

2

1 2

2

0 1

0 −

1

0

0

0

0

2 0

0 0

0 0

0 0

0 0

0

0

0

0

1 2 1 2 0 0 0 0 0

1 2 1 − 2 0 0 0 0 0

0

0

2 0

     1 0  0 0  − 1 0 0 0

 0 0 0 0  0 0 0 0   0 0 0 0  0 0 0 0  1 0 0 0  0 1 0 0  0 0 1 0 0 0 0 1

 0 0 0 0 0 0  0 0 0 0 0 0  1 0 0 0 0 0  0 1 0 0 0 0 0 0 1 0 0 0  0 0 0 1 0 0 0 0 0 0 1 0  0 0 0 0 0 1

ITCH - ELECTRO 2007

Octubre 17-19, Chihuahua, México

From a two level analysis of TF rows, an entirely sub-band is suppressed for all rows. If sub-band 1 is eliminated, 50% of low frequency coefficients reconstruct the image. Value of 75% suppressed coefficients corresponds to eliminate jointly levels one and two. This elimination could be third-level scaled and sub-bands combined in order to decrease image components and increase image compression; however in this case the image distortion is increased. Mathematically, the coefficient suppression can be seen as (4) TFi (sub - band(s)) = 0

The IDWT process is obtained the same way as the DWT but using the transpose of Hl (see equation (3)):

TE is = H Tl * TFiT , (5) i = 1,2,..., m × n / 64 , l = 1,2,3. s

where TE forms the reconstructed matrix from the suppressed coefficients. Then the 64-point rows of

TE is

reconstruct the correspondent sub-images following the same Hilbert fractal curve, generating the new image, where the dimensions are equal to original image.

i = 1,2,..., m × n / 64 where TFi corresponds to the i-th row of TF matrix. At this step, Huffman coding is used. This codification method is based on symbol probability [4] [5], forming new code-words in order to remove coding redundancy, so the zero columns of the TF matrix allow great compressed data due to the 0 symbol repetition. When encoding the symbols, Huffman coding yields the smallest possible number of code symbols per source symbol [2]. The first step in Huffman coding is to create a series of source reduction by ordering the probabilities of the symbols under consideration [2]. The greatest probability symbol is replaced by a single symbol code. The largest coding symbol corresponds to the lowest probability source symbol. Then a “dictionary” is created from the source data. Using this dictionary, a coding vector is created from the original data. An example of Huffman coding dictionary is shown in Table 1.

4. Simulation results A 256 x 256 pixels image is decomposed in 8x8 sub-images and each one is translated to a TEi vector following the described Hilbert fractal curve. In this case i = 1,2,...,1024 , so a 1024x64 TE matrix is then generated. A 2-level Haar Wavelet transform is applied to each of the TE rows, forming the TF matrix, which is the same size as TE. Coefficients of a TFi sub-band must be all entirely suppressed, so it is called sub-band suppression. An example of a vector with suppressed sub-band 1 is plotted in Figure 3.

Table 1. Huffman coding dictionary.

Symbol

A2 A6 A1 A4 A3 A5 A8 A7

Prob 0.4 0.3 0.1 0.06 0.05 0.05 0.02 0.02

Code 0 10 110 1110 11110 111110 1111110 1111111

Figure 3. Vector ssuppression uppression of subsub-band 1.

Huffman coding algorithm is then applied to the TE matrix in order to achieve a good data compression. Multi-symbol Huffman coding [5] would achieve good results too, due to the large rows and columns of zeros in the TF matrix, the probability of symbols composed only by zeros is clearly high. An improvement in data compression is obtained when Huffman coding algorithm is used to encode only the non sub-band suppressed part of TE matrix. The results shown in this paper are based on this sub-band coding only. The inverse wavelet transform is in charge of forming the rest part of TE matrix allocating columns of zeros where necessary.

If it is assumed a data vector A={A5 A1 A2 A6} the correspondent encoding vector is H=[111110110010]. If An is composed by three bits, it is clearly observed that the new vector is not lower than the source vector, but using high dimension data this encoding algorithm yields great compression values. This coding algorithm is applied to a vector which contains the non suppressed sub-band coefficients. The decoding process is obtained directly, and it is done based on the dictionary codes. To reconstruct the image the inverse process consists on decoding dat from the Huffman algorithm, forming again a TF matrix. The Inverse Discrete Wavelet Transform (IDWT) is applied to each TF row.

3

ITCH - ELECTRO 2007

Octubre 17-19, Chihuahua, México

Figure 4. Original image.

Figure 6. Image with subsub-bands 1 & 2 suppressed. Table 2. 2. Numerical results for 256x256 image.

% Coeff 50 25

RC

BPP

3.9 6.8

SNRms

2 1.2

50 27

The mean square signal-to-noise-ratio (SNRms) [2] is obtained by M −1 N −1

SNR ms =

∑∑ fˆ ( x, y )

2

x = 0 y =0

∑∑ [ fˆ ( x, y ) − f ( x, y )]

M −1 N −1

(6)

2

x = 0 y =0

where f ( x, y ) and fˆ ( x, y ) correspond to the original image and resulting image values respectively, x and y are the image coordinates, and M and N are the image matrix dimensions. The BPP (Bits Per Pixel) value is calculated from the compression ratio RC as follows

Figure 5. Image with subsub-band 1 suppressed.

Using an image with clearly distinguished uniform and detailed zones, two cases are presented; suppression of first sub-band, and jointly suppression of first and second sub-bands. The original image is shown in Figure 4. Figures 5 and 6 illustrate the resultant images of the 1, and jointly 1 and 2 suppressed sub-bands respectively. From Figure 6 it is noticeable a great image distortion in more detailed areas. The image contains high changes tone zones, so it generates high frequency components with greater magnitudes in those high slope tones sections (borders of figures). Figures 5 and 6 are reconstructed from 50% and 25% original coefficients respectively. This generates an image with some kind of diffusive border figures. Compression ratio RC and other numerical results for this type of images are shown in Table 2, where BPP denotes Bits per Pixel. A 50% of suppressed coefficients corresponds to sub-band 1 suppressed, and 25% to suppressed jointly sub-bands 1 and 2.

BPP =

8 RC

(7)

Using a greater image (512 x 512) and obtained from a picture (photo), the same cases are presented: suppression of sub-bands 1, and jointly 1 & 2. The original image is shown in Figure 7. Figures 8 and 9 show the resultant images of the mentioned sub-band supression, and Figure 10 shows the error

( fˆ ( x, y ) − f ( x, y )) of

the 25% coefficients reconstructed image. By eliminating sub-band 1, only 50% of the overall original coefficients reconstruct the image. With jointly sub-bands 1 and 2 suppressed, only 25% of the coefficients form the image. However, from Figure 9, it is noticeable a greater image distortion. Numerical results using image from Figure 7 are given in Table 3.

4

ITCH - ELECTRO 2007

Octubre 17-19, Chihuahua, México

Figure 7. Original image.

Figure 9. Image with subsub-band 2 suppressed.

Figure 8. Image with subsub-band 1 suppressed.

Figure 10. Reconstructed image error.

Table 3. 3. Compression and SNRms results.

Im size 256x256 512x512 256x256 512x512

% Coeff 50 50 25 25

RC

2.3 2.3 3.7 3.8

BPP 3.5 3.5 2.1 2.1

As it is observed from figure 10, reconstructed image error is concentrated in figure borders. It is shown that the error presents low values, even with a reconstruction from 25% of the original coefficients. If Lempel-Ziv coding algorithm [2] is used instead of Huffman coding, some variations of it can achieve greater compression ratios. Lossy Lempel-Ziv [1] algorithm allows additional compression ratio but it generates greater image distortion. Also adaptively sub-band suppression can be implemented in order to select the sub-band with lower slopes, and suppress it for each row TFi. This can be based on statistic information measurements, as mean and variance, for each TFi row. Lower image distortion may be achieved, and drawing images could also be compressed by using adaptive methods; however, the computational delay time process is increased. The reported results are obtained by using the Haar mother wavelet, which contains the basis Haar functions, but different image distortion and SNR can

SNRms 241 652 140 363

It is clearly seen that in low size images ( 256 × 256 and lower) SNR values decrease considerably, so distortion values increments by applying this sub-band suppression method. So it is advisable to use the proposed approach in images with high qualities and dimensions. Based on human visual system, using pictures (photos) the resulting images are better than using drawing images due to the lower tone slopes that are involved. However, images with differentiated distributed tone zones (as Figure 4) allow greater compression ratios.

5

ITCH - ELECTRO 2007

Octubre 17-19, Chihuahua, México

be generated by applying other discrete wavelet types like Daubechies, Morlet, or Walsh [3]. Analyzed images are composed by 256 grayscale levels (8 bitsper-pixel resolution), so an analysis for color images can be implemented using this method for each of the RGB component matrixes, achieving good compression ratios and SNR values.

References

5. Conclusions

[3] C. Sidney Burrus, Ramesh A. Gopinath, Introduction to Wavelets and Wavelet Transforms, Prentice Hall, 1998.

[1] Vladimir Crnojevic, Vojin Senk, Zeljen Trpovski, “Lossy Lempel-Ziv Algorithm for Image Compression”, IEEE conference, Serbia and Montenegro, October, 2003. [2] Rafael C. Gonzalez, Richard E. Woods, Digital Image Processing, second Edition, Prentice Hall, 2001.

Based on high frequency coefficient elimination, the use of wavelet transform achieves better results than using other spectral transforms, allowing higher compression ratios and SNR. The 1-D 64-point wavelet transform allows a faster computational performance than those realized by 2-D wavelet transforms [8]. This allows higher dimensions images processing in lower time period. The Hilbert fractal curve is used for subimage 2-D to 1-D decomposition due to the fact that follows a more distributed pattern than a linear scan. For sub-band suppression it achieves lower image distortion. The proposed method can be used for other purposes, such as image de-noising, where the noise elements are located in high frequency bands; and border detection. Wavelet transforms represent a good tool in image processing increasing results quality and performance.

[4] Abraham Lempel, Jacob Ziv, “Compression of TwoDimensional Data”, IEEE Transactions on Information Theory, Vol32, No 1, January 1986. [5] John G. Proakis, Digital Communications, Fourth Edition, Mc Graw Hill, 2001. [6] Sonja Grgic, Kresimir Kers, Mislav Grgic, “Image Compression Using Wavelets”, ISIE’99 IEEE Conference, Bled Slovenia, 1999. [7] Sonja Grgic, Mislav Grgic, and Branka Zovko-Cihlar, “Performance Analysis of Image Compression Using Wavelets”, IEEE Transactions on Industrial Electronics, Vol 48, No 3, June 2001. [8] Hongwei Zhang, Zhengguang Liu, Hougxin Chen, “Low Memory, Low Complexity Line-based Wavelet Image Compression”, IEEE Conference on Signal Processing’04 Proceedings, 2004.

6

an approach to lossy image compression using 1 ... - Semantic Scholar

images are composed by 256 grayscale levels (8 bits- per-pixel resolution), so an analysis for color images can be implemented using this method for each of ...

541KB Sizes 0 Downloads 322 Views

Recommend Documents

an approach to lossy image compression using 1 ... - Semantic Scholar
In this paper, an approach to lossy image compression using 1-D wavelet transforms is proposed. The analyzed image is divided in little sub- images and each one is decomposed in vectors following a fractal Hilbert curve. A Wavelet Transform is thus a

An Adaptive Weighting Approach for Image Color ... - Semantic Scholar
video mobile phones are popular and prevalent. Mobility and ... interpolation (here we call such coalition demosizing). By this .... Left: by AW; center: by. DBW ...

An Adaptive Weighting Approach for Image Color ... - Semantic Scholar
Embedded imaging devices, such as digital cameras and .... enhancement”, Signals, Systems and Computers, 2000, Vol. 2, pp. 1731- ... House (Proposed).

Nonrigid Image Deformation Using Moving ... - Semantic Scholar
500×500). We compare our method to a state-of-the-art method which is modeled by rigid ... Schematic illustration of image deformation. Left: the original image.

Nonrigid Image Deformation Using Moving ... - Semantic Scholar
To illustrate, consider Fig. 1 where we are given an image of Burning. Candle and we aim to deform its flame. To this end, we first choose a set of control points, ...

Image-Based Localization Using Context - Semantic Scholar
[1] Michael Donoser and Dieter Schmalstieg. Discriminative feature-to-point matching in image-based localization. [2] Ben Glocker, Jamie Shotton, Antonio Criminisi, and Shahram. Izadi. Real-time rgb-d camera relocalization via randomized ferns for ke

Compression Progress, Pseudorandomness ... - Semantic Scholar
perbolic discounting of future rewards, humans have an additional preference for rewards ... The unifying theme in all of these activities, it is argued, is the active ...

image compression using deep autoencoder - GitHub
Deep Autoencoder neural network trains on a large set of images to figure out similarities .... 2.1.3 Representing and generalizing nonlinear structure in data .

An Ontology-driven Approach to support Wireless ... - Semantic Scholar
enhance and annotate the raw data with semantic meanings. • domain ontology driven network intelligent problem detection and analysis. • user-friendly visual ...

An Agent-based Approach to Health Care ... - Semantic Scholar
Dept. of Electronic Engineering, Queen Mary & Westfield College, Mile. End Road ... Abbreviated title: Agent-based health care management ... care occurred - a serious concern as time is such a critical factor in care administration). Thirdly ...

Multi-Sentence Compression: Finding Shortest ... - Semantic Scholar
Proceedings of the 23rd International Conference on Computational ... sentence which we call multi-sentence ... tax is not the only way to gauge word or phrase .... Monday. Figure 1: Word graph generated from sentences (1-4) and a possible ...

Using Text-based Web Image Search Results ... - Semantic Scholar
top to mobile computing fosters the needs of new interfaces for web image ... performing mobile web image search is still made in a similar way as in desktop computers, i.e. a simple list or grid of ranked image results is returned to the user.

Using Text-based Web Image Search Results ... - Semantic Scholar
In recent years, the growing number of mobile devices with internet access has ..... 6 The negative extra space is avoided by the zero value in the definition of ...

Characterizing Optimal Rates for Lossy Coding ... - Semantic Scholar
degree, for reasons described in the proof below. Although this result is ..... IEEE Symposium on Foundations of Computer Science, 2003. [11] N. H. Bshouty, Y.

An active feedback framework for image retrieval - Semantic Scholar
Dec 15, 2007 - coupled and can be jointly optimized. ..... space when user searches for cars images. ..... Optimizing search engines using click through data.

using rapd markers - Semantic Scholar
based on this, cluster analysis was done using minimum variance algorithm. Cluster analysis showed two major groups. Each sub-group was characterized ...

i* 1 - Semantic Scholar
labeling for web domains, using label slicing and BiCGStab. Keywords-graph .... the computational costs by the same percentage as the percentage of dropped ...

using rapd markers - Semantic Scholar
RAPD data were used to calculate a Squared Euclidean Distance matrix, and based on this, cluster ... Africa, South-East, Asia, U.S.A, Brazil, Australia and. Turkey. In some ... homogenate was cooled to room temperature and extracted with 5 ...

i* 1 - Semantic Scholar
linear-algebra methods, like stabilized bi-conjugate gradient descent .... W . We can use this property to pre-normalize by first defining λi =σ + β W. { }i* 1. +γ L.

LARGE SCALE NATURAL IMAGE ... - Semantic Scholar
1MOE-MS Key Lab of MCC, University of Science and Technology of China. 2Department of Electrical and Computer Engineering, National University of Singapore. 3Advanced ... million natural image database on different semantic levels defined based on Wo

Domains and image schemas - Semantic Scholar
Despite diÄering theoretical views within cognitive semantics there ...... taxonomic relation: a CIRCLE is a special kind of arc, a 360-degree arc of constant.

Domains and image schemas - Semantic Scholar
Cognitive linguists and cognitive scientists working in related research traditions have ... ``category structure'', which are basic to all cognitive linguistic theories. After briefly ...... Of course, reanalyzing image schemas as image. 20 T. C. ..

Error Correction on a Tree: An Instanton Approach - Semantic Scholar
Nov 5, 2004 - of edges that originate from a node are referred to as its degree. In this Letter we discuss primarily codes with a uniform variable and/or check node degree distribution. Note that relations between the ..... [9] J. S. Yedidia,W. T. Fr

Scalable search-based image annotation - Semantic Scholar
for image dataset with unlimited lexicon, e.g. personal image sets. The probabilistic ... more, instead of mining annotations with SRC, we consider this process as a ... proposed framework, an online image annotation service has been deployed. ... ni