Polarization spectrum image fusion method based on non-subsampled contourlet transformation

Abstract

The invention belongs to the field of image fusion, and particularly relates to a polarization spectrum image fusion method based on non-subsampled contourlet transformation. The method specifically comprises the following steps: S1, collecting polarization spectrum images through a polarization spectrum imaging system, and obtaining spectrum images and light intensity images I of four different linear polarization angles (0 degree, 45 degrees, 90 degrees and 135 degrees); and S2, corresponding Stokes parameters of the image are calculated, and corresponding Q, U and DoLP images are obtained. S3, performing non-subsampled contourlet transform decomposition on the Q image and the I image, averaging low-pass sub-bands, performing preliminary fusion on high-pass sub-bands through an amplitude maximum rule to obtain a polarization feature S, finally performing non-subsampled contourlet transform decomposition on the S image and the intensity image, and fusing the low-pass sub-bands obtained through decomposition by adopting an improved pulse coupling neural network method; and fusing the high-pass sub-bands by adopting a weighted fusion rule based on local mode spectrum characteristics.

Description

Technical Field

[0001] The present invention relates to the technical field of image fusion, and in particular to a polarization spectrum image fusion method based on non-subsampled contourlet transform. Background Art

[0002] In modern military operations, with the continuous advancement of camouflage technology, soldiers, weapons and ammunition and other military supplies are highly similar to the environment, making it difficult for traditional optical target recognition technology to effectively identify camouflaged targets; polarization imaging technology captures the polarization characteristics of light on the surface of an object, enhances the contrast between the object and the background, and provides additional image information; fusing polarization information with spectral information can more comprehensively describe the characteristics of the target object; this multi-dimensional information helps to more accurately identify and distinguish different types of objects, especially in complex backgrounds; the polarization-spectral image fusion method is proposed as a technology for identifying target information in complex backgrounds.

[0003] Previous polarization spectral image fusion methods are prone to spectral distortion due to the different light sources, polarizations and spectra of the previous and next images. In particular, the fused image contains too much redundant information and the image edge information is obviously lost. Therefore, we propose a polarization spectral image fusion method based on non-subsampled contourlet transform to solve the above problem. Summary of the invention

[0004] Technical issues solved

[0005] In view of the deficiencies in the prior art, the present invention provides a polarization spectrum image fusion method based on non-subsampled contourlet transform, which solves the problems raised in the above background technology.

[0006] (II) Technical solution

[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0008] A polarization spectrum image fusion method based on non-subsampled contourlet transform comprises the following steps:

[0009] S1, prepare polarization spectrum image: collect polarization spectrum image through polarization spectrum imaging system, obtain polarization spectrum images and light intensity image I at four different linear polarization angles (0°, 45°, 90° and 135°);

[0010] S2, data preprocessing: calculate the corresponding Stokes parameters of the image and obtain the corresponding Q, U and DoLP images;

[0011] S3, preliminary fusion: The polarization spectrum image and visible light image with the same edge information are decomposed by non-subsampled contourlet transform, the low-pass sub-band is averaged, and the high-pass sub-band is preliminarily fused by the maximum amplitude rule to obtain the polarization feature S;

[0012] S4, image fusion: non-subsampled contourlet transform is used to decompose the polarization features, intensity images and polarization degrees. The low-pass sub-bands obtained by decomposition are fused using an improved pulse coupled neural network method, and the high-pass sub-bands are fused using a weighted fusion rule based on local pattern spectrum features;

[0013] S5, image reconstruction: the image fused in S4 is reconstructed by using non-subsampled contourlet transform to obtain the final fused image;

[0014] S6, image evaluation: The fusion effect of the image is evaluated using subjective forms and objective indicators. The objective indicators include standard deviation, information entropy, and average gradient.

[0015] Furthermore, the method for the four directional parameters and the linear polarization image in the image preprocessing in S2 is to calculate the Stokes vector of the polarization images in four directions (0°, 45°, 90°, 135°) in the polarization data set and to quantitatively describe the intensity image S0 and the polarization image DoLP using the Stokes vector as shown in the following formula:

[0016]

[0017] Where I is the sum of the intensity of the polarized components in the two coordinate axis directions; Q represents the intensity difference image of linear polarization in the 0° and 90° directions; U represents the intensity difference image of linear polarization in the 45° and 135° directions.

[0018] Furthermore, in the S3 preliminary fusion, the non-subsampled contourlet transform is an improved multi-scale and multi-directional image transformation method, which avoids the information loss problem in the traditional contourlet transform by using a non-subsampled pyramid and a non-subsampled directional filter bank;

[0019] The non-subsampled contourlet transform decomposes the image into sub-bands of multiple scales and directions, effectively capturing image details and directional features while maintaining the translation invariance of the image, thereby providing high-quality processing results in image fusion applications; the non-subsampled contourlet transform includes two main steps:

[0020] First, the image is decomposed at multiple scales using a non-subsampled pyramid to generate high-pass subbands and low-pass subbands. Then, the high-pass subband is directional-decomposed using a non-subsampled directional filter bank to divide the frequency domain into multiple subbands, and these subbands are organized through a layered binary tree structure.

[0021] The non-subsampled contourlet transform is mainly divided into the following two steps. First, a non-subsampled pyramid is used for multi-scale decomposition, which is decomposed into two parts: a high-pass subband and a low-pass subband. Then, a non-subsampled directional filter group is used to perform directional decomposition on the high-pass subband. The frequency domain is divided into multiple subbands using a layered binary tree. The number of decomposition layers of the non-subsampled contourlet transform is 2. The first layer performs 4-directional decomposition and the second layer performs 8-directional decomposition.

[0022] Further, in the S4 image fusion, the local spectrum mode is characterized by converting each local area in the image into the frequency domain and extracting texture features therefrom;

[0023] First, for each pixel of the image, a small window around it is selected as the local area. Secondly, Fourier transform is applied to convert these local areas from their original spatial domain to the frequency domain. In the frequency domain, various frequency components can be detected, which reveal the texture information of the local area. Then, amplitude features are extracted from these frequency components and encoded to obtain the representation of the local texture.

[0024] The high-pass subband focuses on the details of the image, such as the edge, texture, and fine structure, emphasizing the details of the image and helping to determine the boundaries and structure of the image; the weights are calculated through the local pattern spectrum, and the steps are as follows:

[0025] (5) Calculate the LSP of each pixel p of S, Q1 and DoLP images respectively s (p), and LSP DoLP (p); (6) For each local mode spectrum feature, calculate E S (p), and E DoLP (p);

[0026] (7) Energy normalization processing:

[0027]

[0028] In the above formula, E S (p), and E DoLP (p) are the local pattern spectrum features of the same pixel point in image S, Q1, and DOLP image, respectively, and W S (P) is related to E S (p) The normalized weight associated with the pixel represents the ratio of the polarization component S at the pixel point p to that at the pixel point s; is with The related normalized weight represents the ratio of the polarization component Q1 at the pixel point p to that at the pixel point 1; W DOLP (P) is related to E DOLP(p) The normalized weight associated with it represents the ratio of the polarization component DOLP at the pixel point p to that at the pixel point .

[0029] (8) The high frequency sub-band fusion formula is:

[0030]

[0031] In the above formula, H S (p), H DoLP (p) represent high-pass subbands at different scales in the same direction, H f (p) is the fused high-pass subband.

[0032] Furthermore, in the S4 image fusion, the high-pass sub-band fusion of the local pattern spectrum features may have a large difference in magnitude. In order to make the weight more stable and easy to compare, the local pattern spectrum features are normalized, and the local pattern spectrum weights of the same position of different images are calculated as follows, and then the high-pass sub-bands are fused. The fusion rule is as follows:

[0033]

[0034] H f (p)=W1(p)·H1(p)+W2(p)·H2(p)

[0035] In the above formula, E1(p) and E2(p) are calculated energies for each local mode spectrum feature; In the above formula, H f (p) is the fused high-frequency sub-band pixel quality, and W1(p) and W2(p) are weights calculated based on the local pattern spectrum features.

[0036] Furthermore, the low-pass subband in the S4 image fusion is fused by combining the improved pulse coupled neural network method. The low-pass subband represents the part of the original image where the grayscale value changes more evenly. The low-pass subband is the area where the grayscale distribution is relatively uniform in the original image, mainly showing the global features and overall structure of the image, including the global information and lower frequency details in the image. The specific steps of the low-pass subband fusion combined with the pulse coupled neural network are as follows:

[0037] (8) Calculate the j-th layer coefficient C of S, Q1 and DoLP images respectively A,j(x,y) , C B,j(x,y) ;

[0038] (9) C A,j(x,y) , C B,j(x,y) Perform normalization processing and input the F channel to activate the pulse coupled neural network;

[0039] (10) Initialize, so that Lj (x,y,0)=U j (x,y,0)=0,θ j (x,y,0)=1, after initialization, the pixel will not be activated, so Y j (x,y,0)=0, the number of pulses generated is T j (x,y,0)=0;

[0040] (11) According to formula (2), L j (x,y,n),U j (x,y,n)θ j (x,y,n),Y j (x,y,n);

[0041] (12) Calculate the number of pulses T j (x,y,n)=T j (x,y,n-1)+Y j (x,y,n);

[0042] (13) In order to achieve better fusion effect, a matching parameter is introduced. The matching degree of two images is defined as follows:

[0043]

[0044] T in the above formula A,j (x,y,n) and T B,j (x,y,n) is in N max The excitation time of the two images in the j-th layer decomposition after iterations; the threshold T in fusion th It can be calculated by the following formula

[0045]

[0046] In the above formula

[0047] (14) The low-frequency subband coefficients can be fused according to the following rules:

[0048] d) When M j ≤T th And T A,j (x,y,N max )≥T B,j (x,y,N max ), so that C F,j (x,y)=C A,j (x,y);

[0049] e) When M j ≤T th And T A,j (x,y,N max)<T B,j (x,y,N max ), so that C F,j (x,y)=C B,j (x,y);

[0050] f) When M j >T th , so that C F,j (x,y)=αC A,j (x,y)+βC B,j (x,y), where

[0051]

[0052] Using C F,j (x,y) is used as the low-pass subband coefficient, and the fused image is obtained after reconstruction.

[0053] (III) Beneficial effects

[0054] Compared with the prior art, the present invention provides a polarization spectrum image fusion method based on non-subsampled contourlet transform, which has the following beneficial effects:

[0055] The present invention aims at the problem that the traditional polarization image fusion method is insufficient in processing the details, nonlinear information, noise and directional information in the image; a more complex multi-scale and multi-directional transformation fusion algorithm is proposed, which is a multi-scale fusion algorithm based on non-subsampled contourlet transform, combined with the characteristics of local pattern spectrum for maintaining local similarity, adaptive weight distribution, good noise resistance and robustness, and an improved pulse coupled neural network algorithm is used in the low-frequency part; the algorithm of the present invention and traditional fusion algorithms such as LRD and PCA evaluate the fusion results subjectively and objectively, and the subjective observation of the comparison between the fused image and the original image shows that the algorithm of the present invention is better than other algorithms. In terms of objective data, the algorithm of the present invention is improved by 5.18%, 7.69%, 30.82% and 10.29% respectively in four evaluation indicators of EN, AG, SNR and SF compared with the experimental optimal fusion algorithm except the present method. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall flow chart of the present invention;

[0057] Figure 2 This is the algorithm fusion flow chart of the present invention;

[0058] Figure 3 It is a schematic diagram of decomposition of non-subsampled contourlet transform of the present invention;

[0059] Figure 4(a): original image; (b): LRD fusion effect; (c): non-subsampled contourlet transform fusion effect; (d): PCA fusion effect; (e): wavelet transform fusion effect; (f): fusion effect of the method in this paper. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] like Figure 1-4 As shown, one embodiment of the present invention proposes a polarization spectrum image fusion method based on non-subsampled contourlet transform, comprising the following steps:

[0062] S1, prepare polarization spectrum image: collect polarization spectrum image through polarization spectrum imaging system, obtain polarization spectrum images and light intensity image I of four different linear polarization angles (0°, 45°, 90° and 135°).

[0063] S2, data preprocessing: calculate the corresponding Stokes parameters of the image and obtain the corresponding Q, U and DoLP images.

[0064] Use I 0° , I 45° , I 90° , I 135° Respectively represent the polarization hyperspectral images obtained at four angles of 0°, 45°, 90°, and 135°; where I represents the sum of the light intensities of the polarization components in the two coordinate axis directions; Q represents the intensity difference of linear polarization in the 0° and 90° directions; U represents the intensity difference of linear polarization in the 45° and 135° directions. The Stokes vector calculation is shown in the following formula:

[0065]

[0066] The value of the degree of linear polarization (DoLP) indicates the proportion of linearly polarized light in the light wave to the total light intensity, which can be expressed as follows:

[0067]

[0068] S3, preliminary fusion: The polarization spectrum image and visible light image with the same edge information are decomposed by non-subsampled contourlet transform, the low-pass sub-band is averaged, and the high-pass sub-band is preliminarily fused by the maximum amplitude rule to obtain the polarization feature S.

[0069] Firstly, the image is decomposed at multiple scales using the non-subsampled pyramid (NSP) to generate high-pass subbands and low-pass subbands. Then, the high-pass subband is directional-decomposed using the non-subsampled directional filter bank (NSDFB) to divide the frequency domain into multiple subbands, and these subbands are organized through a layered binary tree structure. Figure 2 The decomposition process of non-subsampled contourlet transform is shown.

[0070] The non-subsampled contourlet transform is mainly divided into the following two steps. First, NSP is used for multi-scale decomposition, which is decomposed into high-pass and sub-band low-pass sub-bands. Then NSDFB is used to decompose the high-pass sub-band in a directional manner, and the frequency domain is divided into sub-bands using a layered binary tree. The number of decomposition layers of the non-subsampled contourlet transform is 2, with 4 directional decompositions in the first layer and 8 directional decompositions in the second layer.

[0071] S4, image fusion: non-subsampled contourlet transform is used to decompose the polarization features, intensity images and polarization degrees. The low-pass sub-bands obtained by decomposition are fused using an improved pulse coupled neural network method, and the high-pass sub-bands are fused using a weighted fusion rule based on local pattern spectrum features;

[0072] The principle of local pattern spectrum is based on converting each local area in the image to the frequency domain and extracting texture features from it; for each pixel of the image, a small window around it is selected as the local area. Fourier transform or other spectrum transformation methods are applied to transform these local areas from their original spatial domain to the frequency domain. Various frequency components are detected in the frequency domain, which reveal the texture information of the local area; in the local pattern spectrum, each channel represents the relationship between the center point and one of its neighboring points; its formula is:

[0073]

[0074] in, represents the kth channel of the local mode spectrum at point (x, y), f C is a convolutional feature with multiple channels, and (δx k ,δy k ) represents the offset of the kth adjacent point; Φ(·,·) is a function that measures the similarity of two vectors, which will output a scalar; the numerical stability of cosine similarity is selected. Next, features are extracted from these frequency components, such as statistics of amplitude, energy or phase, and these features are encoded to obtain a representation of local texture; finally, these encoded features are mapped back to the central pixel corresponding to each local area in the image, thereby constructing a complete feature image; this feature image retains the local texture distribution and detail information in the original image;

[0075] A pulse coupled neural network is a feedback network composed of a certain number of neurons. Each neuron consists of three parts: a branch tree, a link field, and a pulse generator, which can be described as a mathematical model as follows;

[0076]

[0077] In formula (11), I j and J j is the input of the neuron, Y j is the output of the neuron, U j is the result of internal modulation. The branching tree has two parts, feeding the input F j and link input L j . M kj and W kj are the synaptic gain coefficients of the branch tree and connection field between neuron j and neuron k, respectively, and is the time decay constant, β j is the joint strength, and are the amplification factor and time constant of the threshold integrator, θ j is the dynamic threshold.

[0078] S5, image reconstruction: The image fused in S4 is reconstructed using non-subsampled contourlet transform to obtain the final fused image.

[0079] S6, image evaluation: Evaluate the fusion effect of the image using subjective forms and objective indicators. Objective indicators include standard deviation, information entropy, average gradient, etc.

[0080] To verify the effectiveness of the proposed method, the method was compared with several common polarization fusion methods, including wavelet transform, principal component analysis (PCA), Laplace reconstruction decomposition (LRD) and non-subsampled contourlet transform. Among them, the non-subsampled contourlet transform takes the coefficient with the largest absolute value for the fusion rule of the high-pass sub-band and takes the mean for the low-frequency sub-band.

[0081] Information entropy (EN), standard deviation (STD), average gradient (AG), peak signal-to-noise ratio (PSNR) and spatial frequency (SF) are used as indicators to objectively evaluate the fusion effect. The size of the information entropy value can measure the amount of image information. High information entropy generally means richer detail retention. It is based on the gray value distribution of the image and represents the randomness and uncertainty of different gray levels in the image. The standard deviation reflects the dispersion of the gray value of the image. The higher the value, the higher the contrast of the image and the more dispersed the gray value distribution. The standard deviation can characterize the overall contrast and brightness changes of the image. The average gradient can quantify the sharpness of the image. The higher the gradient value, the higher the sharpness of the image and the more obvious the edges and textures. PSNR is an indicator of image quality. The higher the value, the closer the processed image is to the original image, that is, the better the signal fidelity. Spatial frequency represents the rate of change of gray values ​​in the image and reflects the richness of details and texture complexity of the image. High frequency means that the image contains more details and textures. Different image fusion indicators are shown in Table 1.

[0082] Table 1 Different image fusion indicators

[0083] EN STD AG PSNR SF This method 7.5567 58.7277 74.8020 17.5765 97.8477 LRD Fusion 5.9975 28.7237 48.3215 6.6726 59.0037 NSCT fusion 7.1652 35.5965 69.0479 10.4781 88.0135 PCA fusion 6.3314 59.4546 66.2833 12.1589 87.7759 Wavelet transform fusion 6.0505 30.2075 42.2925 8.1540 50.2012

[0084] exist Figure 3 It can be seen subjectively that the contrast of LRD fusion and wavelet transform fusion is relatively low. The data in Table 1 also confirms that the STD values ​​of the two are much lower than those of the other methods. The shortcoming of PCA fusion is that the fused image has poor ability to retain details, thereby reducing the effective information obtained from the fused image. However, the contrast effect of this fusion method is very good. The evaluation indicators of conventional non-subsampled contourlet transform fusion are relatively balanced, but the difference between the fused image and the original image is large, and the signal fidelity is poor.

[0085] In the objective index evaluation in Table 1, the proposed method is superior to the other four fusion methods except STD. The contrast is 1.24% lower than the PCA fusion method. The four evaluation indicators of EN, AG, SNR and SF are improved by 5.18%, 7.69%, 30.82% and 10.29% respectively compared with the best fusion method except this method. Subjectively, the fusion image of the proposed method has more edge information and texture features, which is closer to the original image and retains more effective information. In summary, the local pattern spectrum feature fusion algorithm based on non-subsampled contourlet transform proposed in this paper is superior to the other four image fusion algorithms.

[0086] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A polarization spectrum image fusion method based on non-subsampled contourlet transform, characterized by: The following steps are included: S1, prepare polarization spectrum image: collect polarization spectrum image through polarization spectrum imaging system, obtain polarization spectrum images and light intensity image I at four different linear polarization angles (0°, 45°, 90° and 135°); S2, data preprocessing: calculate the corresponding Stokes parameters of the image and obtain the corresponding Q, U and DoLP images; S3, preliminary fusion: The polarization spectrum image and visible light image with the same edge information are decomposed by non-subsampled contourlet transform, the low-pass sub-band is averaged, and the high-pass sub-band is preliminarily fused by the maximum amplitude rule to obtain the polarization feature S; S4, image fusion: non-subsampled contourlet transform is used to decompose the polarization features, intensity images and polarization degrees. The low-pass sub-bands obtained by decomposition are fused using an improved pulse coupled neural network method, and the high-pass sub-bands are fused using a weighted fusion rule based on local pattern spectrum features; S5, image reconstruction: the image fused in S4 is reconstructed by using non-subsampled contourlet transform to obtain the final fused image; S6, image evaluation: The fusion effect of the image is evaluated using subjective forms and objective indicators. The objective indicators include standard deviation, information entropy, and average gradient.

2. The polarization spectrum image fusion method based on non-subsampled contourlet transform according to claim 1 is characterized by: The method for the four direction parameters and linear polarization degree image in the image preprocessing in S2 is to calculate the polarization images of the four directions (0°, 45°, 90°, 135°) in the polarization data set and use the Stokes vector to quantitatively describe the intensity image S0 and the polarization degree image DoLP as shown in the following formula: Where I is the sum of the intensity of the polarized components in the two coordinate axis directions; Q represents the intensity difference image of linear polarization in the 0° and 90° directions; U represents the intensity difference image of linear polarization in the 45° and 135° directions.

3. The polarization spectrum image fusion method based on non-subsampled contourlet transform according to claim 1, characterized in that: In the S3 preliminary fusion, the non-subsampled contourlet transform is an improved multi-scale and multi-directional image transformation method, which avoids the information loss problem in the traditional contourlet transform by using a non-subsampled pyramid and a non-subsampled directional filter bank; The non-subsampled contourlet transform decomposes the image into sub-bands of multiple scales and directions, effectively capturing image details and directional features while maintaining the translation invariance of the image, thereby providing high-quality processing results in image fusion applications; the non-subsampled contourlet transform includes two main steps: First, the image is decomposed at multiple scales using a non-subsampled pyramid to generate high-pass subbands and low-pass subbands. Then, the high-pass subband is directional-decomposed using a non-subsampled directional filter bank to divide the frequency domain into multiple subbands, and these subbands are organized through a layered binary tree structure. The non-subsampled contourlet transform is mainly divided into the following two steps. First, a non-subsampled pyramid is used for multi-scale decomposition, which is decomposed into two parts: a high-pass subband and a low-pass subband. Then, a non-subsampled directional filter group is used to perform directional decomposition on the high-pass subband. The frequency domain is divided into multiple subbands using a layered binary tree. The number of decomposition layers of the non-subsampled contourlet transform is 2. The first layer performs 4-directional decomposition and the second layer performs 8-directional decomposition.

4. The polarization spectrum image fusion method based on non-subsampled contourlet transform according to claim 1, characterized in that: In the S4 image fusion, the local spectrum mode is characterized by converting each local area in the image into the frequency domain and extracting texture features from it; First, for each pixel of the image, a small window around it is selected as the local area. Secondly, Fourier transform is applied to convert these local areas from their original spatial domain to the frequency domain. In the frequency domain, various frequency components can be detected, which reveal the texture information of the local area. Then, amplitude features are extracted from these frequency components and encoded to obtain the representation of the local texture. The high-pass subband focuses on the details of the image, such as the edge, texture, and fine structure, emphasizing the details of the image and helping to determine the boundaries and structure of the image; the weights are calculated through the local pattern spectrum, and the steps are as follows: (1) Calculate the LSP of each pixel p of S, Q1 and DoLP images respectively s (p), and LSP DoLP (p); (2) For each local mode spectrum feature, calculate E S (p), and E DoLP (p); (3) Energy normalization processing: In the above formula, E S (p), and E DoLP (p) are the local pattern spectrum features of the same pixel point in image S, Q1, and DOLP image, respectively, and W S (P) is related to E S (p) The normalized weight associated with the pixel represents the ratio of the polarization component S at the pixel point p to that at the pixel point s; is with The related normalized weight represents the ratio of the polarization component Q1 at the pixel point p to that at the pixel point 1; W DOLP (P) is related to E DOLP (p) The normalized weight associated with it represents the ratio of the polarization component DOLP at the pixel point p to that at the pixel point . (4) The high frequency sub-band fusion formula is: In the above formula, H S (p), H DoLP (p) represent high-pass subbands at different scales in the same direction, H f (p) is the fused high-pass subband.

5. The polarization spectrum image fusion method based on non-subsampled contourlet transform according to claim 1, characterized in that: In the S4 image fusion, the high-pass sub-band fusion of the local pattern spectrum features may have a large difference in magnitude. In order to make the weight more stable and easy to compare, the local pattern spectrum features are normalized. The local pattern spectrum weights of the same position in different images are calculated as follows, and then the high-pass sub-bands are fused. The fusion rule is as follows: H f (p)=W1(p)·H1(p)+W2(p)·H2(p) In the above formula, E1(p) and E2(p) are calculated energies for each local mode spectrum feature; In the above formula, H f (p) is the fused high-frequency sub-band pixel quality, and W1(p) and W2(p) are weights calculated based on the local pattern spectrum features.

6. The polarization spectrum image fusion method based on non-subsampled contourlet transform according to claim 1, characterized in that: The low-pass subband in the S4 image fusion is fused by combining the improved pulse coupled neural network method. The low-pass subband represents the part of the original image where the gray value changes more evenly. The low-pass subband is the area where the gray value distribution is relatively uniform in the original image. It mainly shows the global features and overall structure of the image, including the global information and lower frequency details in the image. The specific steps of combining the pulse coupled neural network with the low-pass subband fusion are as follows: (1) Calculate the j-th layer coefficient C of S, Q1 and DoLP images respectively A,j(x,y) , C B,j(x,y) ; (2) C A,j(x,y) , C B,j(x,y) Perform normalization processing and input the F channel to activate the pulse coupled neural network; (3) Initialize, so that L j (x,y,0)=U j (x,y,0)=0,θ j (x,y,0)=1, after initialization, the pixel will not be activated, so Y j (x,y,0)=0, the number of pulses generated is T j (x,y,0)=0; (4) According to formula (2), calculate L j (x,y,n),U j (x,y,n)θ j (x,y,n),Y j (x,y,n); (5) Calculate the number of pulses T j (x,y,n)=T j (x,y,n-1)+Y j (x,y,n); (6) In order to achieve better fusion effect, a matching parameter is introduced. The matching degree of two images is defined as follows: T in the above formula A,j (x,y,n) and T B,j (x,y,n) is in N max The excitation time of the two images in the j-th layer decomposition after iterations; the threshold T in fusion th It can be calculated by the following formula In the above formula (7) The low-frequency subband coefficients can be fused according to the following rules: a) When M j ≤T th And T A,j (x,y,N max )≥T B,j (x,y,N max ), so that C F,j (x,y)=C A,j (x,y); b) When M j ≤T th And T A,j (x,y,N max )<T B,j (x,y,N max ), so that C F,j (x,y)=C B,j (x,y); c) When M j >T th , so that C F,j (x,y)=αC A,j (x,y)+βC B,j (x,y), where Using C F,j (x, y) is used as the low-pass subband coefficient, and the fused image is obtained after reconstruction.

Images