[MATLAB image fusion] particle swarm optimization adaptive multispectral image fusion [including source code phase 004]

One 、 Code run video （ Bili, Bili ）

【Matlab Image fusion 】 Particle swarm optimization adaptive multispectral image fusion 【 Including source code 004 period 】

Two 、 Introduction to image fusion

1 Specific steps of image fusion
(1) Image preprocessing for panchromatic image and multispectral image , Including image filtering 、 Resampling 、 Image registration .
(2) The preprocessed multispectral image fmul Conduct IHS Transformation , Get... Separately fmul-i( brightness )、fmul-h( tonal )、fmul-s( saturation )3 Weight .
(3) For Panchromatic Images fpan And multi spectral image luminance components fmul-i Conduct J layer Contourlet Transformation , Get the following components :
among ,AJ Indicates the low frequency component ;Djk Represents the high-frequency component in each direction on each scale ;j=1,…,J, Represents the decomposition hierarchy ;k=1,…,2lj, It means the first one j Layers in all directions .
(4) Yes Contourlet The low-frequency coefficients and high-frequency coefficients after transformation and decomposition are fused respectively , For the low-frequency coefficients, an adaptive weighted fusion rule based on particle swarm optimization is used , For the high frequency coefficients, the regional structure similarity fusion rule based on particle swarm optimization is adopted , Get the low frequency and high frequency components that meet the fusion requirements .
(5) adopt Contourlet A new luminance component is obtained by inverse transformation fmul-i′, Finally, the new luminance component fmul-i′ And multispectral image components fmul-h And fmul-s Conduct IHS The fused image is obtained by inverse transformation .

2 Low frequency coefficient fusion rule
Contourlet The low frequency coefficients after transform and decomposition reflect the approximate information of the source image , It contains the average features of the source image 、 Spectral information and most of the energy information , Determines the approximate contour of the fused image . The purpose of low frequency coefficient fusion is to properly integrate the characteristic information of panchromatic image on the basis of effectively maintaining the spectral energy information of multispectral image . At present, the commonly used low-frequency coefficient fusion rules are weighted average method 、 The absolute value is greater ( Small ) Method and the larger standard deviation , These methods are intuitive 、 Simple and fast , However, it can not accurately optimize the spectral information of multispectral images and the spatial information of panchromatic images . therefore , This paper presents a low-frequency coefficient fusion rule combined with particle swarm optimization algorithm , Take the linear weighted value as the decision variable to be optimized , The difference between information entropy and relative deviation is taken as the objective fitness function , The optimal weighted fusion coefficient can be found adaptively through the evolutionary iteration of particle swarm optimization algorithm . The specific fusion rules are as follows :

among ,AJfnew-i Represents the low-frequency component after fusion ;w1,w2 Denotes the weighting coefficient ;AJfpan Represents a panchromatic image fpan Low frequency component of ;AJfmul-i Represents the luminance component of a multispectral image fmul-i Low frequency component of .
Low frequency coefficient fusion based on particle swarm optimization algorithm , The difference between information entropy and relative deviation is taken as the objective fitness function , The particle swarm optimization algorithm is used to iteratively find the optimal weighting coefficient when the difference between the information entropy and the relative deviation is the largest w1 and w2. Objective fitness function = Information entropy (E)- Relative deviation (RD), The information entropy is the information entropy of the fused image , It is an evaluation based on the amount of information ; The relative deviation is the relative deviation between the fused image and the original multispectral image , It is based on the evaluation of spectral performance . To search for the optimal weighting coefficient , The parameters of particle swarm optimization algorithm are set as follows :c1 by 0.01,c2 by 0.02, Inertia weight ω by 0.02, The population size is 20, The maximum number of iterations is 600.

3 High frequency coefficient fusion rule
The high frequency coefficient reflects the details of the source image , It mainly contains edges 、 texture 、 Feature information such as bright lines and areas , But a single image pixel can not well represent these characteristic information , Therefore, it is necessary to comprehensively consider and analyze the relationship between the corresponding pixels of multi-source images , It is characterized and reflected by multiple pixels of this regional feature . The purpose of high frequency coefficient fusion is to preserve more spatial details of the original panchromatic image while maintaining better spectral characteristics . In this paper, references are given [19] The high-frequency coefficients are fused by the fusion rules based on the similarity of regional structure , On this basis, the global optimization ability of particle swarm optimization algorithm is used to find the optimal threshold of regional structure similarity p To fuse the high-frequency coefficients , The fusion rules are as follows .

(1) Particle swarm optimization algorithm is used to find the optimal threshold of structural similarity between high-frequency coefficients p, In order to determine the high-frequency coefficient fusion . In the algorithm proposed in this paper , To fuse the structural similarity between the image and panchromatic image (SS) As the objective fitness function , The particle swarm optimization algorithm is used to iteratively find the threshold value when the structural similarity is the largest p, Then the fusion rules based on the similarity of regional structure are used for fusion . The parameters of particle swarm optimization algorithm are set as follows :c1 by 0.01,c2 by 0.02, Inertia weight ω by 0.6, The population size is 20, The maximum number of iterations is 300.
(2) Window operation for high-frequency directional subbands , Calculate the structural similarity of their corresponding regions , And record the value of similarity .
(3) If the similarity is less than p, The principle of maximum standard deviation is adopted for fusion , Formula for :
If the similarity is greater than p, Use the following weighted fusion rules , Formula for :
among ,std Represents the standard deviation of coefficients in the neighborhood window ;j Represents the decomposition hierarchy ,j=1,…,J;k It means the first one j Layers in all directions ,k=1,…,2lj;SSIMjk(fpan,fmul-i) by fpan and fmul-i The structural similarity of the corresponding region ;E1jk and E2jk Respectively fpan、fmul The weight of the corresponding area .

When the structural similarity of the corresponding regions of the two source images is less than the optimal threshold p when , It shows that the image correlation is small , The fusion method with the largest regional variance can increase the details of the fused image as much as possible ; When the structural similarity of the corresponding regions of the two source images is greater than the threshold p when , It shows that the two images are highly correlated and similar , The weighted average fusion method can preserve more common regional structure features of the source image . Two images X、Y The structure similarity of is defined as follows :

among ,mX and mY They represent images respectively X、Y The average of ;σX2 and σY2 They represent images respectively X、Y The variance of ;βXY Represents an image X、Y The covariance ;L(X,Y)、C(X,Y) and S(X,Y) They represent images respectively X、Y The brightness of 、 Contrast 、 Structure comparison ;C1、C2、C3 Is a small constant , To avoid instability when the denominator is zero .

3、 ... and 、matlab Edition and references

1 matlab edition
2019b

2 reference
[1] Guzhipeng , He Xinguang .Contourlet A remote sensing image fusion method coupled with transform and particle swarm optimization [J]. Computer science . 2016,43(S2)

3 remarks
This part of the introduction is taken from the Internet , For reference only , If infringement , Contact deletion