A FAS INENSIY-HUE-SAURAION FUSION APPROACH VIA PRINCIPAL COMPONEN ANALYSIS FOR IKONOS IMAGERY S. Minhayenud, S. Chitwong and F. Cheeasuit Department of Instrumentation Engineering, Faculty of Engineering King Mongkut s Institute of echnology Ladkrabang Bangkok, 0520 HAILAND kcsakrey@kmitl.ac.th, kcfusak@kmitl.ac.th ABSRAC o enhance spatial information of low resolution multi-spectral (RGB) image, the intensity-hue-saturation (IHS) approach is perfectly used to fuse the low resolution RGB image and the high resolution panchromatic (Pan) image by replacing intensity component with the high-resolution Pan image. Disadantage of the mentioned approach is that color of a fused image is changed because the saturation component is changed or spectral of the low resolution RGB image and the high resolution Pan image is different, that is, spectral information of the fused RGB image is distorted. his problem is important for applying the fused image for classification. o sole this problem, in this paper, we employ the principal component analysis (PCA) transformation to extract information from the low resolution RGB image. In procedure of fusion method, the first principal component is used to adjust brightness of the high resolution Pan image. he intensity component from IHS transformation is replaced by the adjusted brightness high-resolution Pan image. he experimental results by using IKONOS imagery show that the proposed approach is better performance than the original IHS methods by improing spectral distortion and still correlating to the Pan image. INRODUCION Recently, multi-sensor image fusion has been proposed by many researchers based on IHS approach. One of the main disadantages of one is that the spectral of the fused image is distorted. he mentioned problem decrease capability of GIS application to use the fused RGB image for classification and interpretation. he last researching report (Choi, M.,2006) proposed the method to reduce the spectral distortion by using parameter to adjust the intensity of intensity component from RGB-IHS transformation and the Pan image but the parameter not hae the closed form. hen, in this paper, as mentioned concept to reduce spectral distortion based on the IHS approach by which the intensity component and the Pan image must hae the corresponding intensity. Since intensity component in IHS space come from the RGB image, the information used to match brightness of the Pan image to correspond with intensity of the all RGB image must then hae the information of the RGB image more and more possible. One of method to concentrate the all information of the RGB image to single image is principal component analysis and also the first principal component which contains the most information is used to match brightness of them. As mentioned method, it is simple and low computational processing and also can sole the main problem of the conentional IHS approach. his paper is organized as follows. In Section 2 and 3, principal component transformation and fast intensity-huesaturation are briefly described; in Section 4 fusion method and experimental results are described, respectiely. he conclusion is gien in Section 5. PRINCIPAL COMPONEN RANSFORMAION o concentrate the multi-band data to single band, the principal component transformation can be accomplished by the following steps:. Calculate the coariance matrix by C X
k CX = ( Xi M)( X i M), k i= () where X is a gien N-dimensional ariable, M is the mean ector and k is the number of pixels. he principal component or Y i can be denoted by Y = a X + a X + K+ a X i i 2i 2 Ni i i λ 0 0 K 0 0 λ2 0 0 K CX = 0 0 λ3 K 0 M M M O 0 0 0 0 K λ N λ > λ > K > λ. N, (2) Y = a X. (3) All transformed component can be written by Y= A X. (4) where A is the eigenector matrix. he coariance matrix of new Y coariance is obtained by C = AC A, (5) Y X where C is a diagonal matrix, consisting of eigenalue of C, that is X X, (6) where 2 N 2. Project the original image onto the eigenector. he principal component images of each eigenalue can be obtained by the projection of the original images onto the corresponding eigenector. By applying the PCA process to the IKONOS multispectral images as RGB images which are shown in Figure 2, the three principal component images are shown in the Figure 3. able indicates the eigenalaues and percentage of the information of the original RGB images in each of principal component. We see that the first principal component contains information of original RGB image more than 95%. able Information of original RGB images in PCs Index PC PC2 PC3 Eigenalues 8 3.902 0 8 0.47 0 % of RGB 95.836 3.636 0.5503 8 0.00224 0 INENSIY-HUE-SAURAION RANSFORMAION he well-known fusion technique is Intensity-Hue-Saturation transformation (IHS). he intensity component is replaced by the high resolution Pan image. he procedure of transformation is as following. Step. Upsample (resize) the low spatial resolution RGB image to the size of the high spatial resolution Pan image. Step 2. ransform RGB into HIS as equation (6),
3 3 3 I R 2 2 2 = 6 6 6 2 B Step 3. he intensity component I is replaced by the high resolution Pan image. Step 4. Back transform IHS into R G B as equation (7), where R,, 0 R Pan G = B 2 G (6) 2 0 G B, I,, and represent corresponding alues in the original RGB image. R, G, and B are 2 corresponding alues in the fused image. By rewriting (7), a computationally efficient method can be expressed as where R I + (Pan I) G = B 2 2 0 I + δ R+ δ = G δ = + 2 B+ δ 2 0 δ = Pan I (9) Equation (8) states that the fused image [ R, G, B ] can be easily obtained from the original image [ R, GB, ] simply by using addition operation only. hat is, the IHS fusion can be implemented efficiently by this procedure. (7) (8) FUSION MEHOD EXPERIMENAL RESULS o oercome the disadantage of fusion method using IHS approach, that is, the spectral of the fused image is destroyed. In this paper, we propose the fusion method based on the IHS approach as many preious research (Audicana, M. G., Saleta, J. L., Catalan, R. G., and Garcia, R., 2004; Choi, M.,2006; u,.-m., Huang, P. S., Hung, C.- L., and Chang, C.-P., 2004), but the simple technique is represented by adjusting brightness of the high resolution Pan image shown in Figure 4(a) to correspond with that of the first principal component from the original RGB image.
Figure shows the fusing scheme consisting of up-sampling process which is resize of low resolution image into high resolution image, RGB to IHS to RGB transformation, principal component transformation, and brightness matching. he intensity component from RGB-IHS transformation is then replaced by the adjusted brightness high resolution Pan image shown in Figure 4(b) and back transform to R G B image as fused RGB image shown in Figure 5(b), while Figure 5(a) shows the fused RGB image from directly replacing the intensity component with the high resolution Pan image. In our experiment, we use the IKONOS image of 28-by-28 sizes for low resolution RGB image and 52-by- 52 sizes for high resolution Pan image. he principal components can be deried from the original low resolution RGB image using principal component transformation as described in preious section. We selected the first principal component PC, which had approximately 95% of information in the original RGB image. he brightness of Pan image is adjusted to correspond with that of intensity component which is the PC image, namely I. Intensity component I from the RGB-IHS transformation is replaced by I image. Finally, fused RGB image can be deried from I HS to R G B transformation shown in Figure 5(b). o confirm the performance of our fusion method corresponding with reducing spectral distortion and high resolution RGB image, we basically use correlation coefficient CC and standard deiation SD, that is, CC alue between each of the fused RGB images and the Pan image reach to while SD alues of difference image between the original RGB images and the fused RGB images reach to 0. Both CC and SD in able 2 indicate that the CC alues of our fusion method are more than that of IHS fusion method and SD alues are less than. able 2 he numerical performance index Index Image IHS Our IHS CC R -Pan 0.9370 0.9500 G -Pan 0.9405 0.9532 B -Pan 0.8483 0.882 SD R-R 7.0483 5.4742 G-G 6.6567 5.552 B-B 6.5240 5.0489 CONCLUSIONS In this paper, we present the fusion method as simple technique by fusing image from RGB image and the Pan image at which its brightness is adjusted to correspond to the first principal component deried from the original RGB image and also containing information more than 95%. he intensity component from RGB-IHS transformation is replaced by the adjusted brightness Pan image. he fused RGB image from our fusion method show that spectral distortion can reduce by indicating SD alues and it is the high resolution multi-spectral image by indicating CC alues. herefore, the fused RGB image of I HS-RGB will be useful for terrain feature classification and interpretation. REFERENCES Audicana, M. G., Saleta, J. L., Catalan, R. G., and Garcia, R. (2004). Fusion of multispectral and panchromatic images using improed IHS and PCA mergers based on waelet decomposition. IEEE rans. Geoscience and Remote Sensing., 42(6):29-299. Choi, M. (2006). A New intensity-hue-saturation fusion approach to image fusion with a tradeoff parameter. IEEE rans. Geoscience and Remote Sensing., 44(6):672-682. u,.-m., Huang, P. S., Hung, C.-L., and Chang, C.-P. (2004). A fast intensity-hue-saturation fusion technique with spectral adjustment for IKONOS imagery. IEEE Letters Geoscience and Remote Sensing, (4):309-32.
Figure. he fusing scheme. (a) (b) (c) (d) Figure 2. he original RGB IKONOS image (a) Red image (b) Green image (c) Blue image (d) RGB image.
PC PC2 PC3 Figure 3. he three principal component image. (a) (b) Figure 4. he high resolution Pan image (a) Original Pan image (b) the Adjusted brightness Pan image.
(a) (b) Figure 5. he fused RGB image (a) intensity component replaced by Pan Image (b) replaced by the adjusted brightness Pan image.