Image Processing Reference

In-Depth Information

suitable for shape extraction of the objects in the fused image due to their property

of removing the image details without adding any gray level bias.

Burt and Kolczynski [20] have proposed a generalized multisensor fusion using

the gradient pyramid where the process of generation of the image pyramid has

been referred to as the
pyramid transform
. Using the basis functions of gradient-of-

Gaussian pattern, they apply the pyramid transform to both of the input images. Each

of these basis functions is derived from the single prototype function via shifting,

scaling, and rotation. This process decomposes the image up to three levels of suc-

cessive approximations along four orientations. To combine the information across

multiple decompositions, authors define two fusion rules. At the locations where the

source images are similar, the fusion is achieved by averaging two images, while if

the images are significantly different, the fusion rule selects the feature pattern with

maximum saliency and copies it for the fused image.

Liu et al. have demonstrated the use of a steerable pyramid for fusion of remote

sensing images [107]. The steerable pyramid is a multi-scale and multi-orientation

decomposition with translation and rotation invariant sub-bands [167]. The low

frequency or the coarsest approximation is fused based on the magnitude of the

images at the corresponding locations. The fusion rule for the high frequency details

is derived from the strength of each of the orientations which gives the directional

information at that location. Authors refer to this selection rule as the absolute value

maximum selection (AVMS).

Piella [135] has proposed a region-based technique in a multi-resolution frame-

work as an extension of the pixel-based technique. This work provides a generalized

structure for multi-resolution fusion techniques. The input images are first segmented

which is a preparatory step toward the actual fusion. The author uses the term
activity

measure
which captures the saliency in the image. The other quantity is the match

measure which quantifies the similarity between the corresponding coefficients of

the transformed images. This structure encompasses most of the pixel-based and

region-based multi-resolution techniques, and also can be considered as the basis for

the development of new ones.

For an efficient fusion, one needs to extract the salient features from multi-scale

image decompositions. The wavelet transform has proved to be a highly popular

tool for fusion. We assume that the readers are familiar with the fundamentals of the

wavelet transform. For details on wavelets, one may refer to the dedicated texts by

Daubechies [46], Vaidyanathan [178], Mallat [111], etc. The advantages offered by

wavelets along with their theoretical background can also be found in [5, 124].

Li et al. have proposed a discrete wavelet transform (DWT)-based technique for

fusion as wavelets offer distinct advantages such as orthogonality, compactness,

and directional information [104]. Their technique is claimed to be superior to the

Laplacian pyramid-based techniques as it does not produce any visible artifacts in the

fused image as opposed to the later. Similar to the multi-resolution fusion approaches

discussed earlier, the basic principle behind the wavelet-based fusion techniques is as

followsâ€”the set of input images is first decomposed into different multi-resolution

coefficients that preserve image information. These coefficients are appropriately

combined at each level to obtain new coefficients of the resultant image. This image is