Image Fusion Based on the Wavelet Transform

Transcription

1 Journal of Information & Computational Science 5: 3 (2008) Available at http: Image Fusion Based on the Wavelet Transform Kaicheng Yin a, Weidong Yu a Textile materials and technology laboratory, Donghua University, Shanghai ,P.R.China b Department of college of garment and art design, JiaXing University,Zhejiang ,P.R.China a, b,* Abstract In this paper, we focus on the image fusion based on the wavelet transform and focuses on image fusion between multispectral images and panchromatic images. In order to enhance the result image's spatial resolution with preserving information, research on algorithms is developed in this paper. Some experiment results are given. Moreover, some kinds of existing algorithms are introduced and summarized in detail. Keywords: Image Fusion; Remote Sensing Image; Wavelet Transform 1 Introduction Multi-sensor system has been widely used in various fields.with the continuous increase sensors of system,it make the system available information to show diversity and the dramatic increase in the amount of information,previous information processing method has been unable to meet this new situation,and it didn t not to apply the new problems,which the use of multi-sensor combination brought about in a position.information Fusion is a new method to meet this demand.information fusion is processing and integrated,which is multi-level, multi-faceted and multi-level from a number of sensors obtained the multi-source information,so as to gain a richer, more accurate, more reliable, useful information.the information of The signals from multiple sensors provided is redundant and complementary, and information fusion system can maximize access to objective information or a complete description of scene. Since the early 1980s, multi-sensor image fusion has triggered a worldwide study upsurge,which have broad application prospects in computer vision, remote sensing, robotics, medical image processing, and military fields. Although people have developed the various sensors, forward-looking infrared, laser imaging radar, synthetic aperture radar, etc,to obtain * Corresponding author Weidong Yu. addresses: hyykc@mail.dhu.edu.cn (Kaicheng Yin), wdyu@dhu.edu.cn (Weidong Yu) / Copyright 2008 Binary Information Press May, 2008

2 1380 K. Yin et al. / Journal of Information & Computational Science 5:3 (2008) high-quality image, their performance greatly dependent on the use of the environment, and any independent image sensor function has its limitations.the image information obtained from Integration of multiple sensors can improve the shortcomings of a single sensor.multi-source Image Fusion is the use different images obtained from various imaging sensors or different imaging modalities of the same imaging sensors or different images obtained from different imaging time. The image of complementary integrated information and redundant information will obtain more comprehensive and accurate description of the image for further treatment or observation.through the integration of them, we can expand the scope of sensors and increase confidence and improve system reliability and maintainability and lower right of a single sensor.performance requirements can be more accurate and more reliable to obtain a more comprehensive vision or Scene description.in recent years, with the development of sensor technology and computing capacity, multi-sensor image fusion technology applications was more widely. 2 Methods of Remote Sensing Transform Integration With rapid development of remote sensing technology, satellite remote sensing systems of the multi-platform, multi-temporal, multi-spectral and high-resolution for complex application-oriented background rush a large number,and it was used in a wide range of applications for environmental observation and surface area of plant classification and so on.to be able to feature a more accurate classification and more precise positioning, a higher spatial resolution and higher spectral resolution remote sensing images are required. To achieve this goal there are two methods. One method is using remote sensing image sensor of a higher spatial resolution and higher spectral resolution, but the methods was higher cost, and will be limitations of the actual level of technical and technological. Another method is to use image processing technology, Image fusion of high spatial resolution panchromatic images and high spectral resolution and the integration of multi-spectral images was made.to obtain image fusion of both high spatial resolution and high spectral resolution, the method can save a lot of production costs and research funding.in remote sensing applications, multi-spectral imaging sensor will obtain a number of remote sensing images of spectrum, but the low-resolution image clarity is poor. Panchromatic imaging sensors will obtain high-resolution panchromatic remote sensing image, but it has not a spectral information. Image fusion of high-resolution and multi-spectral imaging will obtain a clear high-resolution and multi-spectral image. 2.1 Method of IHS transform fusion At present, people conducted extensive research and made many algorithms about high spatial resolution panchromatic images and the high spectral resolution multi-spectral image fusion.ihs color space conversion based on the integration of remote sensing images is a fusion algorithm in the application of remote sensing.in algorithm, the three-band of multi-spectral images are mapped RGB (Red-Green-Blue, red - green - three-color blue) color space, and then made color space conversion from RGB color space to the HIS. integration is a process of using

3 K. Yin et al. / Journal of Information & Computational Science 5:3 (2008) high-resolution panchromatic images instead of I Component (luminance component) of IHS color space, and then obtained fused image from IHS color space to RGB color space conversion.ihs color space conversion of remote sensing image fusion algorithm is simple and can be used to retain the high-resolution panchromatic images of the ground resolution information, but it will introduce more severe spectral (color) distortion.ihs transform fusion method is the most commonly method of multi-source integration inremote sensing images.ihs transform is the color space conversion,and IHS transform have wide application of the flexible and practical advantages,and it become a mature standard method of image fusion. 2.2 Method of PCA transform fusion PCA (Principal Component Analysis) Transform also called Principal Component Analysis, or K L transform, which is the multi-dimensional orthogonal linear transformation based on the statistical characteristics.pca is a methods that make more components integrated few components through drop dimensional technology.pca transform widely used in image compression, random noise signal and image rotation and removal.pca transform is that many of the original necessarily related weight regrouped into a new group of mutual non-integrated component to replace the original component.the original P components is a linear combination as a new component though mathematical treatment.the first linear combination is F1.In order to the linear combination with a unique, it was requirements that the largest variance is F1 in all linear combination.if var (F1) is more bigger, and then the more information included.if F1 cann t stand for the original P-component information, the second principal component F2 would consider selecting.information of F1 has requested not in F2, and this means cov (F1, F2) = 0, and the rest may be deduced by analogy until the full expression of the original information.in fact, X Vector Image PCA Transform is problems of the desire for image covariance matrix R eigenvector. When the PCA transform image imposed the image from transform will be the best approximation of the original image. For different purposes of the fusion,pca transform of the traditional fusion algorithm is different on image components choice, and then steps and fusion algorithm is not the same.the multi-spectral images principal component transform is three steps in the traditional algorithms. The first principal component was the replacement of high-resolution images, and the final it was inverse transform back again, and obtained finally fused images. In order to maintain consistency of image pixel, before the replacement high-resolution images with the multi-spectral image brightness components need revised in histogram. 2.3 Method of wavelet transform fusion Recently, remote sensing image fusion algorithm based on multi-scale decomposition has been widely used in remote sensing image fusion, and remote sensing image fusion algorithms based on discrete wavelet transform (DWT) is typical.discrete wavelet transform image fusion algorithms, bring forward by Zhou, J., is used for remote sensing of the Landsat TM images and multi-spectral panchromatic high-resolution SPOT Image Fusion.

4 1382 K. Yin et al. / Journal of Information & Computational Science 5:3 (2008) First, discrete wavelet transform was made from multi-spectral Landsat TM images of the band and SPOT panchromatic images, and then SPOT Image DWT coefficients of the low frequency band was displacede by TM images band discrete wavelet transform coefficient of the low-frequency subbands. Final,discrete inverse wavelet transform was made to get image fusion from the TM-band multi-spectral images and SPOT panchromatic images.remote sensing image fusion algorithm of the discrete wavelet transform is superior to IHS color space conversion and principal component analysis and remote sensing image fusion based Brovey Transform algorithm. 3 Wavelet Transform and Remote Sensing Image Fusion 3.1 Wavelet Transform The concept of Wavelet was bring forward by the French geophysicist Morlet in the 1984.In analysis of time-frequency seismic wave of local,he hopes to use adaptive Transform for the high-frequency with narrow window and the low-frequency with broaden window.fourier transform is difficult to meet this requirement.he then cited the Gaussian modulated cosine function, and it was extendable by a translation function of the group, which was later called "Morlet Wavelet."In 1986, Jafferd, Lemavie,Meyer and Mallat bring forward that orthogonal wavelet-based structure can be incorporated into a unified frame.they introduce the concept of multi-resolution analysis,and they unified specific wavelet bringing forward by predecessors, and they gives a general orthogonal wavelet methods of multiresolution analysis of the structure.with inspired by Brut and Adelson s tower decomposition algorithm,mallat bring forward the rapid decomposition and reconstruction algorithms of the wavelet transform, and now it was called Mallat algorithm.wavelet Transform has the following advantages in Image Fusion: (1)Multiresolution decomposition provided a the image information of different resolution, and the majority energy was concentrated in the low-frequency after transform. So we can carry through the corresponding information integration. (2) Wavelet decomposition and reconstruction algorithm is recycled, which is easy in hardware implementation. (3)It was Facilitate with parallel processing, parallel implementation. 3.2 Images of two-dimensional the wavelet transform Image fusion was made byusing discrete wavelet transform, arithmetic of a binary Mallat discrete wavelet image decomposition was generally using. Mallat two-dimensional image decomposition algorithm can be expressed as a mn, = h h a lk, D 1 mn, = h g a lk, ( ) ( ) ( ) ( ) j 1 l 2m k 2n j j 1 l 2m k 2n j

5 K. Yin et al. / Journal of Information & Computational Science 5:3 (2008) (1) (, ) = (, ) (, ) = (, ) D m n g h a l k D m n g g a l k 2 3 j 1 l 2m k 2n j j 1 l 2m k 2n j Mallat two-dimensional image reconstruction algorithm can be expressed as: (, ) = ( ) 1 ( ) 2 ( ) , , , (, ) a mn h h a lk h g D lk g h D lk g g D lk (2) j m l n k j m l n k j m l n k j m l n k j Here, h stand for analysis of the low-pass filters, and g stand for high-analysis filter,and h ~ stand for the low-pass filter integrated,and g ~ stand for high-integrated filters.usually image can be seen as two-dimensional discrete signal, and two-dimensional discrete wavelet transform is a powerful tool for the image processing.through wavelet transform, the image sequence was available in different spatial resolution,and it was very easy to further analysis and understanding.in the basis of MRA theory, S. Mallat proposed discrete wavelet transform algorithm by subband structure,that is fast wavelet transform algorithm, which unified sub-band filter and wavelet transform calculations. 3.3 Technology of image fusion Image discrete wavelet transform is the horizontal and vertical direction to repeat high - and low-pass filtering of the original image and sampling, and the original image multiresolution denotation was obtained.if the original image f(x, y) as AO f, AO f can be decomposed to under J 2 (J Z + ), which means J 2 and 3J high frequency resolution. AO f can be denotation with a low-frequency resolution J 2 J 2 J j 1 J 2 J j 1 J 2 J j 1 A J f under 2 ( A f,( D f)),( D f),( D f) ) (3) For the Orthogonal of Wavelets, Image wavelet decomposition process will not have redundant data,and wavelet decomposition process is that the signal was decomposition to independent of each other frequency channel of airspace, so that we can facilitate analysis signals frequency domain in various frequency bands.when using Wavelet transform based on Mallat algorithm in the high-resolution, multi-spectral image fusion,(assuming the high-resolution and multi-spectral images with geometric registration), Multi-spectral images will be conversion to IHS color space, and the hue and saturation is maintain the same ( ' H = H, S ' = S ). Information of high-resolution map and multi-spectral images in the strength of IHS space component was integration, and the spectral intensity I' was obtained by integration. Therefore, the inverse transform of I ', H, S was into RGB space, and a high spatial resolution multi-spectral image was obtained.image fusion process based on pixel-level are given as Figure 1. Through separately wavelet transform,a rigorous registration in the two images will obtained a series of sub-band data, and fusion was made in two image in the corresponding location on the basis of subband certain criteria,and then image reconstruction integration was made from the wavelet inverse

6 1384 K. Yin et al. / Journal of Information & Computational Science 5:3 (2008) transform Fig 1 Image fusion diagram based on wavelet transform 4 Experiment and Analysis To validate the algorithm performance, the two groups remote sensing image was using fusion experiments.image in Experiment 1 is Canada Fredericton urban remote sensing imagery taken by commercial satellite IKONOS.IKONOS satellite may shoot multi-spectral remote sensing images of 4-meter ground resolution and panchromatic high-resolution images of 1-meter ground resolution.the remote sensing images in experiment 2 is the London region's 10 meter resolution panchromatic SPOT Image (Figure 3 (a)) and 30-meter resolution Landsat TM multispectral images (Figure 3 (b)).the purpose of integration of remote sensing images is to integrated multi-spectral images of spectral information and high-resolution panchromatic image information. Fig 2 experiment 1 of remote sensing image fusion Fig 3 experiment 2 of remote sensing image fusion The experiment results show that DWT fusion effect was significantly better than HIS transformation, and we can see that IHS color space conversion of remote sensing image fusion algorithm is simple and can be used to retain the high-resolution panchromatic images of the ground resolution information, but will introduce more severe spectral (color) distortion, and DWT overcome this shortcoming.

7 K. Yin et al. / Journal of Information & Computational Science 5:3 (2008) Conclusion we focus on the image fusion based on the wavelet transform and focuses on image fusion between multispectral images and panchromatic images In this paper. In order to enhance the result image's spatial resolution with preserving information, research on algorithms is developed in this paper. We experiment two fusion algorithm. Based on previous work,and through detailed theoretical analysis and experimental verification, Remote sensing image fusion method of IHS transform and the wavelet transform are bring forward,and we research wavelet decomposition of the wavelet function choice and effect of wavelet decomposition scale integration, and system summarize the results of the image fusion method. References [1] MALLAT S. A theory of multiresolution signal decomposition : the wavelet t ransform [J]. IEEE Transactionson Pattern Analysis and Machine Intelligence, 1989, 11(7) : [2] EHL ERS M. Multisensor image fusion techniques in remote sensing [J]. ISPRS Journal of Photogrammetry andremote Sensing, 1991, 46(2): [3] Kingsbury N. The dual2tree comp lex wavelet transform: a newtechnique for shift invariance and directional filters [A ]. In: IEEE Digital Signal ProcessingWorkshop [ C ], Vancouver, BC, Canada,1998, 86: 86~89. [4] Nan Wang, Shuqing Zhao, Jianxing Guo. Imagefusion between SAR image and TM image using IHS and activity2level measured wavelet packet t ransform [J ].Journal of Institute of Surveying and Mapping, 2004, 21(4): [5] Rui Chao, Ke Zhang, Yanjun L i. An image fusion algorithm using wavelet t ransform [J ]. Acta Ectctronica Ainica,2004,32( 5): [6] Jianhong Huang, Qingnian Wang, Yanfen Luo.Autoregressive power spectrum based on wavelet packetf requencyshift algorithm and its application in the extraction of fault features of rolling bearing [J ]. Machinery Design &Manufacture,2005(7):