Image Fusion (v)--gradient pyramid

Image fusion algorithm based on gradient pyramid (Gradient pyramid,gp) decomposition. GP is also a multi-scale decomposition algorithm based on Gaussian pyramid. The GP representation of the image can be obtained by the operation of the gradient operator on each layer of the Gaussian pyramid. GP each layer decomposition image contains horizontal, vertical and two diagonal four directions of detail information, can better extract the image edge information, improve stability and noise resistance. The directional gradient tower decomposition provides a good way to provide the direction edge and detail information of the image.

1, the principle of elaboration

(1) Get Gauss pyramid (as above)

(2) Gradient-type decomposition can be obtained for each decomposition layer (except the highest layer) of the Gaussian pyramid of the image, respectively, by gradient-directional filtering:

Here • For convolution operations, DL K represents the K-direction gradient tower image of the L-layer, GL is the image of the L-layer image of the Gaussian pyramid, and DK represents the K-directional gradient filter operator, defined as:

Through D1, D2, D3, D4 to the Gaussian pyramid of the various layers of the directional gradient filter, on each decomposition layer (except the highest level) can be included in the horizontal, vertical and two diagonal direction of detail information of 4 decomposition images. The gradient tower decomposition of the visible image is not only multi-scale and multiresolution decomposition, but also consists of each decomposition layer (except the highest layer) which contains 4 directions of detail information respectively.

The difference here is that each layer is independent and does not need to involve the upper sampling results on the previous layer. The corresponding layer of GL and 3*3 of the core of the convolution, after adding the value of GL to take the corresponding direction of the value, you can generate the corresponding direction of the coefficient.

(3) Refactoring

After each layer of pyramid image is fused separately, it is necessary to reconstruct the original image by Gradient pyramid, to introduce the FSD Laplace pyramid as the intermediate result, to convert the gradient pyramid to Laplace Pyramid, and then reconstruct the original image by the Laplace pyramid, and the construction process is as follows:

1, the direction gradient pyramid is converted to the direction of the Laplace pyramid (FSD type) filter-subtract-decimate. The first L-layer image of the FSD-type pyramid is LL,

2, transform the FSD Laplace pyramid image into the Laplace pyramid image.

Note I is not a unit matrix, just the middle one element is 1. (not understand)

3, the original image is reconstructed by the Laplace pyramid to enlarge the value of the GL interpolation, so that the enlarged image size is the same size as the GL-1. This is the same as the previous one (Pyrup).

2. Fusion Application

Using region-based fusion rules, the basic idea of region-based fusion method is: In order to determine the pixel of fused image, we should consider the corresponding pixels in the fused source image, and consider the local domain of the fused pixels in the fusion processing of a decomposition layer image. That is to compare the source image of a certain aspect of the characteristics, so that the dynamic selection of the feature prominent source image composition fusion results.

The gradient is a vector that points toward the direction of the maximum value of the local edge in the direction of the edges of the normals, and the edge direction of the image is always orthogonal (perpendicular). Therefore, gradient-based filters, also known as edge operators. After filtering the image by the gradient filter, the change of gray level between adjacent points is highlighted, and the aim of enhancing edge is achieved. Taking the sum of the gray value of each point of the region as the characteristic quantity, when the fusion of the source image decomposition layer, the value of the characteristic from which region is large, the gray value of the center pixel of the region is used as the pixel gray value of the location on the decomposition layer of the fusion image. This will be a good way to extract the image edge information.

the basic steps of fusion are:

1. The gradient pyramid of image is established for each source image by gradient tower.

2, the decomposition layer of the image gradient pyramid is fused separately, the different decomposition layer, the different direction detail image uses the different fusion operator to carry on the fusion processing, finally obtains the fusion image gradient pyramid.

3, to the fusion of the resulting gradient pyramid inverse tower transformation (that is, image reconstruction), the resulting reconstructed image is fused image for the fusion rule can be selected based on the region information, can also be a simple method to take the maximum value.

?

Image Fusion (v)--gradient pyramid