All copy rights Reserved by NATCOMM , Christ Knowledge City, Mannoor, India. Published by IJRCCT ( Page 48

Transcription

1 A Survey on Image Fusion Techniques Bincy Bavachan, Dr.Prem Krishnan Student, Dean PG Studies Department of Computer Science and Engineering Christ Knowledge City, Ernakulam, India Abstract Image fusion is the process by which input images are fused in order to increase the quality. The input images must be the images of the same scene with different quality measures. The quality of the output image will be better than any of the input images. The main methods of the image fusion involve simple image fusion, pyramid based image fusion and wavelet based image fusion. Image fusion methods can be used in areas of medical imaging, remote sensing, entertainment etc. This paper focuses on the different image fusion methods evolved so far and their comparison based on the quality of the output image. It also discuss about the different quality metrics that can be used to assess the quality of the output image. Keyword s Image Fusion; Frequency Domain; Complex Wavelet Tran sform; DT-CWT I. INT RODUCTION scene. Multimodal image fusion fuse images taken fro m different modalities of the same scene. It is widely used in medical imaging. In multi focus fusion, the input images will be focusing in different scenes. Some images may focus on the foreground and some on the background. In multi sensor fusion, the input images will be taken fro m different s ns ors. This method is wide ly used in remote sensing application where the sensors lack quality. The researches in image fusion have found out that fusion can be done both in spatial domain and in spectral domain. Since spatial do main representation This report is organized as follows : Sect ion II presents brief description of spatial based image fusion methods, Section III presents spectral based image fusion methods, Section IV compare the various methods and finally, conclusion is presented in Section V. An image is worth more than ten thousand words. So image processing is of high importance. An image can be represented both in s partial domain and in spectral domain. Image Fusion [1] is the process of fusing perfectly registered input images which can represent both spatial and spectral information. It retains important information and produces an image which is better for human and computer perception for further processing Since spatial representation is s imp le, spatial fusion methods are also simple. Spectral representation can clearly define the edge features of the image. So the spectral based image fusion methods produce better quality output image rather than spatial methods. Simple Image Fusion methods come under the category of spatial methods whereas Pyramid as well as Wavelet trans form methods come under spectral methods. Different types of image fusion [3] are Multi view Fusion, Multimodal Fusion, Multi focus Fusion, Multi Sensor Fusion and Single Sensor Fusion. Multi view Fusion fuses the images obtained from multiple cameras. These images will be taken fro m different views of the same II. SP AT IAL DOMAIN BASED FUSION Spatial do main methods operate directly on pixels. Pixels of the image will be represented in a matrix format. A. Averaging Method [8] Average of pixel P(i,j) of a ll the input images are taken and assigned to the corresponding pixel of the output image. It is the simplest of all the methods and is very easy to understand and implement. Good and bad aspects of an image are minimized arriving at a mean value. B. Select Min/Max Method Here the Max/Min pixel values among all the pixel values of the input images are taken and assigned to output image. The advantage of s elect ma x method is that no compromise is made over the good information available whereas s elect min method is us ed in cas e of darker images where darkness relates to noise. C. Principal Component Analysis Method Published by IJRCCT ( Page 48

2 PCA method [9] enables us to compres s the data without informat ion los s. Method of PCA is as follows :- Get the data. (pixel information) Subtract mean value. Calculate the covariance matrix. Calculate eigen vectors and eigen values of covariance mat rix. Chosing components from the eigen vectors and producing a feature vector. Feature Vector = (e ig1 eig 2 eig 3 eign ) Deriving the new datas et. New Datas et = Feature Vector * Data Adjus t New Datas et of all the input images are formulated and the sum of the pixe l va lues forms the fus ed image matrix. In PCA method, the res ultant fus ed image will be of high s patial quality. D. IHS Transform Fusion [5] It is bas ed on three factors : Hue(H) represents the average wavelength of the light, Intens ity(i) repres ents the brightness of the image and Saturation(S) refers to the purity of the color. Brovey transform fusion [4], also known as color normalized fusion is a method to fus e images fro m different s ens ors. The resultant image can be used for visual interpretation. This method doesn t make any change in the spectral feature of pixels. Rather it trans forms lu minance information to panchromatic images. III. SP ECT RAL DOMAIN BASED FUSION Spectral representation of image is based on the frequency characteristics of the image. Since spatial fusion methods suffer fro m spatial degradation, spectral image fusion methods came into existence. Spectral Image Fusion methods are of two types : Pyramid based fusion methods and Wavelet based fusion methods. A. Pyramid Based Fusion Methods A pyramid image consists of a s et of low pas s filters through which the image pass through. An image at a particular level will be half the size of the same image in the latter level. Pyra mid trans forms of the input images are formed, they are combined and taking the inverse pyramid forms the fus ed image. A pyramid trans form [3]is accomplished by the following three s teps : Decomposition Formation of the initia l image for recompos ition Reco mpos ition Deco mpos ition is the process by which a pyramid is generated at each level of the fus ion process. The follo wing s teps are performed m number of times, if m is the predefined depth of fus ion. Fig 1: IHS Image Fus ion This method is most widely used in commercial applications like re mote sensing. In this method, the RGB color bands are converted into I,H and S components and the inverse trans formation provided by S and H gives the fused image. E. Brovey Transform Low pas s filtering: The input images are filtered by us ing the predefined filters according to the pyramidal method. Pyra mid is formed fro m the filtered input images us ing Burt s method or Lis Method. The input images are decimated to half their s ize, which would act as the input image matrices for the next level of decompos ition. Input images are merged after decompos ition which is the input to recompos ition. The recompos Published by IJRCCT ( Page 49

3 ition is the process by which the final image is created fro m the pyramids formed as a res ult of decompos ition. It is done by performing the following s teps m nu mber of t imes. The input image is undecimated to the level of recompos ition. The undecimated matrix is filtered with the trans pos e of the filter vector us ed in t he decompos ition process The filtered matrix is then me rged with the pyramid formed at the respective level of decompos ition. The newly formed image matrix would act as the input to the next level of recompos ition. The fus ed image be the res ultant image at the fina l level of reco mpos ition process. Diffe rent pyramidal methods differ in the filters they us e for low pas s filtering. 1) Filter Subtract Decimate Pyramid Deco mpos ition phas e filters the input image by the low pass filter, W= [1/16, 4/16, 6/16, 4/16, 1/16]. The low pass filtered input images are subtracted in order to form the pyramid. Now the pyramid is decimated by halving the number of rows and colu mns. Reco mpos ition phase undecimates the pyramid by duplicating the number of rows and columns. Again the pyramid is fed into a low pass filter 2*W. Matrix addit ion of the pyramid at the corres ponding levels form the final pyra mid. 5) Morphological Pyramid In order to avoid the effect of noise, two stages of image filtering are done image opening and image closing. Image opening is performed by combining image erosion after image dilation. Image closing is performed by combining image dilation after erosion. B. Discrete Wavelet Based Fusion Methods A discrete wavelet trans form [6, 7] is a wavelet trans form in which the wavelets are discretely sampled. It captures both frequency and time information and s o it can repres ent even the minute feature of an image. 2) Laplacian Pyramid Laplac ian pyra mid is s imilar to FSD e xcept that the low pass filtering is done by 2*W. 3) Ratio Pyramid It is similar to FSD, but rather than subtraction in the decompos ition phas e, pixe l wis e ratio is calcu lated. 4) Gradient Pyramid The difference between FSD and this method is that a low pas s filter, V = [1/ 2, 1/4, 2/4] is us ed along with the low pas s filter W. In addition to this four directional filters are us ed. Fig 2: 2D DWT for an image Here the signal is decomposed by using a low pass filter, g(n) and a high pas s filter, h(n). Published by IJRCCT ( Page 50

4 Fig 3: Decomposition Discrete Wavelet Trans form DT-CWT [2,8] is an advanced version of DWT which overcome the disadvantages of DWT like directional selectivity and lack of shift invariance. Here fusion is performed fro m the DT -CWT decomposed structure. DT - CWT structure is nothing but the decomposition of an image into two filter banks. The methodology of DT-CWT based fusion is as follows. This series of decomposition is represented as a tree known as filter bank. Fig 4: Filter Bank 1) DWT with Haar Based Fusion Haar Wavelet is the firs t wavelet and the scaling function is described as follows. The input images are converted to their Haar wavelet form, integrated as such and the reverse process will give the fused image. Fig 5: DT-CWT Bas ed Fus ion 2) DWT with Daubechies Based Fusion This category of DWT us es recurrence relation to generate the finer dis crete s amples. The filters cons idered are IV. COMP ARISON OF DIFFERENT FUSION H1 = [ ] H2 = [ ] G1 = [2 4 2] 3) DT- C WT Based Fusion Published by IJRCCT ( Page 51

5 V. REFERENCES [1] Hongbo Wu and Yanqiu Xing, Pixel-based image fusion using wavelet transform for spot and ETM+ image in PIC, IEEE international conference, [2] Kings bury, N. G. 1998a. The dual tree complex wavelet transform: a new technique for shift invariance and directional filters, proc. 8th IEEE DSP Works hop,bryce Canyon, UT, USA, paper no. 86 [3] Zhang Zhong, Investigations on Image Fusion, PhD Thes is, Univers ity of Lehigh, USA. May [4] As hok Mandhare, Pixel Level Image Fuison using Brovey Trans form,ija REEIE,June [5] Firou z Abdullah Al-Was s ai, N.V. Ka lyankar,, The IHS Transformations Based Image Fusion. [6]Shivs ubraman i Kris hnamoorthy, So man K. P, Implementation andcomparative Study of Image Fusion Algorithms, International Journal of Co mputer Applications, Vo l. 19, no. 2, Nov V. CONCLUSION AND FUTURE WORK This paper discussed about different fusion methods emerged s o far. Out of this the DT -CWT method is found to be better when checked with subjective as well as objective quality metrics. Also all the spectral methods are found to be free from spatial distortion. Image registration, if employed to input images before fusion will improve the quality. Image registration is a method used to make the images into a common format even if the images are captured in different cameras, in different lighting and in different angle. [7] C. Sydney, Burrus Ra mes h, A. Gopinath and Haitao Guo, Introduction to wavelets and wavelets transforms A primer, Prentice Ha ll,1998. [8] Sruthy S, Dr.Latha Para mes hwaran, Image Fusion Technique using DT-C WT, IEEE Trans action, [9] Linds ay I Smith, A Tutorial on Princ ipal Co mponent Analys is c453/s tudnent_tut orials /principa l_components.pdf Fusion of images by DT-CWT method after performing IHS trans format ion of the input images will improve the quality of the image fusion algorithm as IHS trans form method separates the Different components of images like intensity, saturation and wavelength. Published by IJRCCT ( Page 52