Image Processing Reference

In-Depth Information

distortion in the result [61, 202]. The PCA-based methods introduce less color dis-

tortion, but affect spectral responses of the multispectral data [61]. This spectral

distortion is caused due to the mismatch of overlap between the spectral responses

of the multispectral image, and the bandwidth of the pan image [61].

We have already discussed the utility of the multi-resolution based methodologies

for fusion. This has also been used for pan-sharpening. The pan image is decomposed

into a set of low-resolution images yielding a pyramidal structure. The low-resolution

multispectral bands replace the pan image at an appropriate level of resolution from

the pyramid. The reverse wavelet transform is then employed on each of the mul-

tispectral bands to produce the corresponding output. A non-orthogonal undeci-

mated multi-resolution decomposition known as the “à trous” wavelet transform is

most commonly employed by several pan-sharpening techniques [1, 4, 39, 149].

Aiazzi et al. have proposed a context driven thresholding of correlation coefficients

between the images to be fused in the wavelet domain in order to avoid injection

of undesired spatial details [1]. Zhang and Hong have integrated the IHS- and the

wavelet-based techniques to reduce the color distortion [202]. In their technique,

the multispectral bands are first projected onto the IHS color space prior to their

wavelet decomposition. After the component substitution, the inverse wavelet trans-

form followed by the inverse IHS transform generates the resultant sharpened image.

In [66], the IHS transform of the multispectral image has been resampled to the size

of the pan image. These images and the pan image are then decomposed using the

Daubechies wavelet where the detail coefficients of the pan image are injected into

the corresponding intensity component.

Contourlets have been known for better directional representation than wavelets,

and capturing the geometrical structure of the objects [50]. Shah et al. have proposed

the use of contourlets along with an adaptive PCA which preserves the spectral

information [162]. Since the curvelets are more suitable for edge representation,

they are also well suited for pan-sharpening. In [39, 40], curvelets have been used to

extract the spatial details from the pan image.

Alparone et al. have proposed to extract the texture details from the pan image

which are used to modulate the intensity of the multispectral image bands [4]. This

technique is different from several others as it does not employ direct combination

of image components. Joshi et al. compute a set of autoregressive (AR) parameters

by modeling the spatial correlation of multispectral bands which are assumed to

be the same for the pan image due to spectral correlation among them [83]. These

parameters are then used in the process of regularization employed to combine the

spectral characteristics of themultispectral image and the pan image. The generalized

intensity component has been modeled as the weighted linear combination of the

multispectral bands in [2]. These weights have been computed as the regression

coefficients between the multispectral bands and the spatially degraded version of

the pan image. The pan-sharpening has been carried out using the Gram-Schmidt

spectral sharpening technique. Garzelli et al. have proposed a procedure to obtain a

pan-sharpened image which minimizes the squared error between the multispectral

data and the fused image [62]. Their technique provides the optimal results in the

sense of themean squared error (MSE). Moeller et al. have developed awavelet-based