CN105718957A - Polarized SAR image classification method based on nonsubsampled contourlet convolutional neural network - Google Patents

Abstract

The invention discloses a polarimetric SAR image classification method based on a non-subsampled contourlet convolutional neural network, which mainly solves the problem that it is difficult to avoid the influence of coherent speckle noise and the problem of low classification accuracy in the prior art. The implementation steps are: treat the classification Denoise the polarimetric SAR image, and perform Pauli decomposition on the denoised polarization scattering matrix S; combine the image features obtained by Pauli decomposition into a feature matrix F, and normalize it, denoted as F1; for each Take the 22 × 22 blocks around F1 to get the block-based feature matrix F2; select the training data set and test data set from F2; construct a non-subsampling contourlet convolutional neural network to train the training data set; use A non-subsampled contourlet convolutional neural network is trained to classify the test dataset. The invention improves the expression ability and classification accuracy of the polarimetric SAR image features, and can be used for target recognition.

Description

Polarimetric SAR image classification method based on non-subsampled contourlet convolutional neural network

technical field

The invention belongs to the technical field of image processing, and in particular relates to a polarimetric SAR image classification method, which can be used for target recognition.

Background technique

Polarization SAR is a high-resolution active microwave remote sensing imaging radar, which has the advantages of all-weather, all-time, high resolution, side-view imaging, etc., and can obtain richer information on targets. The purpose of polarimetric SAR image classification is to use the polarization measurement data obtained by airborne or spaceborne polarimetric SAR sensors to determine the category to which each pixel belongs. It has a wide range of applications in agriculture, forestry, military, geology, hydrology, and ocean research and application value.

The commonly used polarimetric SAR image classification method is based on pixels, that is, only the characteristics of each pixel are used for classification. Although these methods can better preserve the pixel-level details in the image, due to the influence of coherent speckle, there is an error between the measured value of a single pixel and the real value, and it is difficult to avoid the existence of many isolated pixels and small areas in the classification map. , increasing the classification difficulty.

Existing target feature extraction methods based on scattering characteristics in polarimetric SAR images include Cloude decomposition, Freeman decomposition, etc.

In 1997, Cloude et al. proposed Cloude decomposition to divide the H/α plane, and convert each pixel into the category of the corresponding area through the two eigenvalues of H and α that characterize the polarization data. One defect of H/α classification is that the division of regions is too arbitrary. When the same class of data is distributed on the boundary of two or more classes, the performance of the classifier will deteriorate. Another shortcoming is that when the data in the same area When several different features coexist, they cannot be effectively distinguished;

In 2004, Lee et al. proposed a feature extraction method based on Freeman decomposition. This method can maintain various polarization scattering characteristics, but the classification results are easily affected by the performance of Freeman decomposition. For polarization data of different bands, the algorithm poor universality.

These feature extraction methods do not take into account the multi-scale and multi-resolution characteristics of polarimetric SAR images, and it is difficult to obtain high classification accuracy for polarimetric SAR images with complex backgrounds.

Contents of the invention

The purpose of the present invention is to solve the above problems and propose a polarimetric SAR image classification method based on non-subsampled contourlet convolutional neural network to obtain image features with multi-scale and multi-resolution characteristics and improve classification accuracy.

The idea of the present invention is: based on the convolutional neural network to process the image block features, and by introducing non-subsampling contourlet transformation into the network, the expression ability of the polarimetric SAR image features is effectively improved, and the implementation scheme includes the following:

(1) Denoise the polarimetric SAR image to be classified, and obtain the polarimetric scattering matrix S after filtering the polarimetric SAR image;

(2) Perform Pauli decomposition on the filtered polarization scattering matrix S, and use the values of odd scattering, even scattering and volume scattering obtained by Pauli decomposition as the image features of the polarimetric SAR image;

(3) Combine the image features obtained by Pauli decomposition into a pixel-based feature matrix F of the polarimetric SAR image, each pixel corresponds to a 3-dimensional Pauli decomposition feature, and normalize the element values in F to [0, 1], denoted as F1;

(4) Take a 22×22 block around F1 for each pixel to obtain a block-based feature matrix F2, that is, each pixel corresponds to three 22×22 blocks;

(5) Select training data set and test data set from block-based feature matrix F2:

(5a) Divide the polarimetric SAR image ground objects into 15 categories, randomly select N marked pixels from each category as training samples D1, and the remaining marked pixels as test samples T1, N takes 300~ an integer between 700;

(5b) Use the Canny operator to extract the edge points of the polarimetric SAR image, add the edge points extracted by the Canny operator to the training sample D1, that is, increase the training samples with higher confidence, and obtain the updated training data set D and test data set T;

(6) Construct a non-subsampled contourlet convolutional neural network:

(6a) Select an 8-layer convolutional neural network consisting of input layer → convolutional layer → pooling layer → convolutional layer → pooling layer → fully connected layer → fully connected layer → softmax classifier, and determine the convolutional neural network The filter size of the network and the feature maps of each layer;

(6b) Replace the second layer of convolutional layer in the convolutional neural network with a non-subsampling contourlet transformation layer to obtain a non-subsampling contourlet convolutional neural network;

(7) Training the training data set with a non-subsampled contourlet convolutional neural network;

(8) Use the trained non-subsampled contourlet convolutional neural network to classify the test data set, and obtain the pixel category of each pixel in the polarimetric SAR image test data set.

Compared with the prior art, the present invention has the following advantages:

1. The present invention combines pixel spatial correlation to extract image block features, weakens the influence of coherent speckles, and thus improves classification accuracy.

2. Since the present invention adopts non-subsampling contourlet convolutional neural network, and introduces non-subsampling contourlet transformation into the convolutional neural network to obtain image features with multi-scale and multi-resolution characteristics, it can better approximate the original image , improving the classification accuracy.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Fig. 2 is the pseudo-color figure after denoising of image to be classified among the present invention;

Fig. 3 is the manual marking diagram of the image to be classified in the present invention;

Fig. 4 is a diagram of classification results of images to be classified using the present invention.

detailed description

Below in conjunction with accompanying drawing, implementation steps and experimental effects of the present invention are described in further detail:

With reference to Fig. 1, the concrete realization steps of the present invention are as follows:

Step 1, denoise the polarimetric SAR image to be classified.

Commonly used polarization SAR image denoising methods include mean filtering, median filtering, local filtering, refined polarization LEE filtering, etc. The present invention uses the refined polarization LEE filtering method, and the specific steps are as follows:

(1a) Set the sliding window of refined polarization LEE filtering, the size of the sliding window is 5×5 pixels;

(1b) Roam the sliding window from left to right and from top to bottom on the pixels of the input polarimetric SAR image, and move the sliding window from left to right and from top to bottom according to the pixel space position at each roaming step Divide into 9 sub-windows in turn, the size of each sub-window is 3×3 pixels, and there is overlap between the sub-windows;

(1c) average the pixel values at the corresponding positions of each sub-window, and form a mean window of 3×3 pixels with the obtained mean;

(1d) Select the gradient templates in the four directions of horizontal, vertical, 45 degrees and 135 degrees, weight the mean window with the four templates, calculate the absolute value of the obtained weighted results, and select the largest of all absolute values value, the direction corresponding to the maximum value is taken as the edge direction;

(1e) Take two left and right sub-windows in the direction of the edge of the central window from the nine sub-windows, calculate the mean value of all pixel values in the two sub-windows, and subtract the mean value of all pixel values in the center window from the two mean values obtained, Use the sub-window corresponding to the smaller absolute value in the mean difference as the direction window, wherein the center window refers to the 3×3 sub-window at the center of the 5×5 window;

(1f) According to the formula <1>, the weight of the refined polarization LEE filter is obtained:

$\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ></span>$

Among them, b represents the weight of the refined polarization LEE filter, var(y) represents the variance value of the polarization SAR total power image pixels in the direction window, y represents the pixel of the polarization SAR total power image in the direction window, and p represents the direction The mean value of all pixels in the polarimetric SAR total power image in the window, Represents the variance value of the speckle noise of the input polarimetric SAR image;

(1g) According to formula <2>, the polarization coherence matrix T of the central pixel of the filtered polarimetric SAR image is obtained:

T=w+b(z-w), <2>

Among them, w represents the mean value of the polarization coherence matrix of the polarimetric SAR image pixel in the direction window, b represents the weight of the refined LEE filter, and z represents the polarization coherence matrix of the central pixel of the polarimetric SAR image;

(1h) According to formula <3>, the scattering component _SHH of horizontal emission and horizontal reception, the scattering component S _VV of vertical emission and vertical reception, and the scattering component _SHV of horizontal emission and vertical reception can be obtained:

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; V</span>}}{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ {\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; two</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; V</span>}}{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ {\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; 33</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; V</span>}}{<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; ></span>$

Wherein, T ₁₁ , T ₂₂ , and T ₃₃ are elements on the diagonal of the polarization coherence matrix T.

The pseudo-color image of the image to be classified after denoising is shown in Figure 2.

Step 2: Perform Pauli decomposition on the filtered polarimetric scattering matrix S, and use the values of odd scattering, even scattering, and volume scattering obtained from Pauli decomposition as the image features of the polarimetric SAR image.

(2a) Define the basic scattering matrix, called the Pauli basis: {S _a , S _b , S _c , S _d }, the formula is as follows:

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}}}\left[\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right]\\ {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}}}\left[\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right]\\ {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}}}\left[\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; 1</span>}\\ \mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right]\\ {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}}}\left[\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 0</span>}& <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; i</span>}\\ \mathrm{<span\; class="notranslate">\; i</span>}& \mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right]\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; ></span>$

in $\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right]$ stands for odd scattering, $\left[\begin{array}{cc}1& 0\\ 0& -1\end{array}\right]$ represents even scattering, $\left[\begin{array}{cc}0& 1\\ 1& 0\end{array}\right]$ represents volume scattering, $\left[\begin{array}{cc}0& -i\\ i& 0\end{array}\right]$ Indicates the type of feature that does not exist, so the value of d is 0;

(2b) According to the Pauli basis defined by formula <4>, the expression of polarization scattering matrix S is obtained:

$\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; h</span>}}& {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; V</span>}}\\ {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; h</span>}}& {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; V</span>}}\end{array}\right]<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; s</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; wxya</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; ></span>$

Among them, a corresponds to the value of odd scattering, b corresponds to the value of even scattering, c indicates the value of volume scattering, and d indicates the value of the scattering component corresponding to the type of non-existing ground objects;

(2c) Solve formula <5> to obtain the scattering values a, b, c, d, and express them as vectors as follows:

$\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&lsqb;</span>\begin{array}{cccc}\mathrm{<span\; class="notranslate">\; a</span>}& \mathrm{<span\; class="notranslate">\; b</span>}& \mathrm{<span\; class="notranslate">\; c</span>}& \mathrm{<span\; class="notranslate">\; d</span>}\end{array}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\left(\begin{array}{cccc}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; V</span>}}& {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; V</span>}}& {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; V</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; h</span>}}& \mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; V</span>}}<span\; class="notranslate">\; )</span>\end{array}\right)}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; ></span>$

When the reciprocity condition S _HV ＝ S _VH is satisfied, formula <6> is simplified as:

$\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \&lsqb;</span>\begin{array}{ccc}\mathrm{<span\; class="notranslate">\; a</span>}& \mathrm{<span\; class="notranslate">\; b</span>}& \mathrm{<span\; class="notranslate">\; c</span>}\end{array}<span\; class="notranslate">\; \&rsqb;</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}}}\left(\begin{array}{ccc}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; V</span>}}& {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; V</span>}\mathrm{<span\; class="notranslate">\; V</span>}}& \mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; V</span>}}\end{array}\right)<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; ></span>$

Substitute _SHH , S _VV , and _SHV obtained from formula <3> into formula <7> to obtain the polarization characteristic K.

Step 3, combine the image features obtained by Pauli decomposition into a feature matrix F, and normalize it.

Construct a feature matrix F, the size of the matrix is set to M1×M2×3, and assign the values of odd scattering, even scattering, and volume scattering obtained by Pauli decomposition to the feature matrix F, where M1 is the length of the image to be classified, and M2 is the width of the image to be classified;

To normalize the feature matrix F, the feature linear scaling method is used, that is, the maximum value max(F) of the feature matrix F is first obtained; then each element in the feature matrix F is divided by the maximum value max(F) to obtain Normalized feature matrix F1.

Step 4, take a 22×22 block around F1 for each pixel after normalization, and obtain a block-based feature matrix F2, that is, each pixel corresponds to three 22×22 blocks, and the size of the feature matrix F2 is 22 ×22×(M1×M2)×3.

Step 5, select a training dataset and a testing dataset from the block-based feature matrix F2.

(5a) Divide the polarimetric SAR image ground objects into 15 categories, randomly select N marked pixels from each category as training samples D1, and the remaining marked pixels as test samples T1, N takes 300~ Integer between 700;

(5b) Use the Canny operator to extract the edge points of the polarimetric SAR image, add the edge points extracted by the Canny operator to the training sample D1, that is, increase the training samples with higher confidence, and obtain the updated training data set D and test Data set T.

Step 6. Construct a non-subsampling contourlet convolutional neural network.

(6a) Select an 8-layer convolutional neural network consisting of input layer → convolutional layer → pooling layer → convolutional layer → pooling layer → fully connected layer → fully connected layer → softmax classifier, and determine the convolutional neural network The filter size of the network and the feature maps of each layer;

(6b) Replace the second convolutional layer in the convolutional neural network with a non-subsampled contourlet transformation layer to obtain a non-subsampled contourlet convolutional neural network, and obtain the following 8-layer structure:

Input layer→non-subsampling contourlet transform layer→pooling layer→convolution layer→pooling layer→full connection layer→full connection layer→softmax classifier;

The parameters of each layer are:

Layer 1 input layer: output feature map = 3;

The second layer of non-subsampling contourlet transformation layer: output feature map = 12;

The third layer of pooling layer: downsampling scale = 2;

The fourth layer of convolutional layer: output feature map = 20, filter size = 4;

Layer 5 pooling layer: downsampling scale = 2;

Layer 6 fully connected layer: output feature map = 100;

Layer 7 fully connected layer: output feature map = 64;

Layer 8 softmax classifier: output feature map=15.

Step 7, use the non-subsampled contourlet convolutional neural network to train the training data set.

The feature matrix of the training data set is used as the input of the non-subsampled contourlet convolutional neural network, and the network output layer is the corresponding predicted class label. By solving the error between the predicted class label and the manually marked correct class label, and calculating the error Perform backpropagation to optimize the weight of the non-subsampled contourlet convolutional neural network. The error backpropagation method of the present invention is the same as that of the convolutional neural network, and the correct class label of the artificial mark is shown in Figure 3.

Step 8: Use the trained non-subsampled contourlet convolutional neural network to classify the test data set, and obtain the pixel category of each pixel in the polarimetric SAR image test data set.

Effect of the present invention can be further illustrated by following simulation experiments:

1. Simulation conditions:

The simulation experiment uses the full polarization data of the L-band Flevoland area in the Netherlands from the AIRSAR system of the NASA/JPL laboratory, and the image size obtained based on Pauli decomposition is 750×1024 pixels.

The hardware platform is: Intel(R) Xeon(R) CPUE5-2620, 2.00GHz*18, and the memory is 64G.

The software platform is: MATLAB_2014a.

2. Simulation content and results:

Carry out experiments under the above-mentioned simulation conditions with the method of the present invention, namely randomly select 700 marked pixels from each category of the polarimetric SAR data as training samples, and the remaining marked pixels as test samples, the training data set Accounting for 6% of the total number of samples, the classification results shown in Figure 4 are obtained. It can be seen from Figure 4 that except for a few misclassified pixels, the regional consistency of the classification results is good, and the outline is very clear.

Then reduce the training samples successively, so that the training data set accounts for 5%, 4%, and 3% of the total number of samples, and compare the accuracy of the test data set and the model training time of the present invention with the convolutional neural network, and the results are as shown in Table 1:

Table 1

As can be seen from Table 1, when the present invention accounts for 6%, 5%, 4%, and 3% of the total number of samples in the training data set, the accuracy of the test data set is higher than that of the convolutional neural network, and the required model training time is shorter; When the training data set accounts for 3% of the total number of samples, the present invention can obtain a classification accuracy of 92%, but the classification accuracy of the convolutional neural network does not converge.

In summary, the present invention effectively improves the expression ability of polarimetric SAR image features, improves classification accuracy, and reduces model training time by introducing non-subsampling contourlet transform into convolutional neural network. When the sample size is small, the advantage is obvious.

Claims (5)

1. a Classification of Polarimetric SAR Image method for non-down sampling contourlet convolutional neural networks, including:

(1) Polarimetric SAR Image to be sorted is carried out denoising, obtain the filtered polarization scattering matrix S of Polarimetric SAR Image；

(2) filtered polarization scattering matrix S is carried out Pauli decomposition, Pauli is decomposed obtain odd scattering, even scattering, volume scattering value as the characteristics of image of Polarimetric SAR Image；

(3) Pauli is decomposed the eigenmatrix F based on pixel of the polarization SAR image of image characteristic combination obtained, the corresponding 3 dimension Pauli characteristics of decomposition of each pixel, and the element value in F is normalized between [0,1], it is denoted as F1；

(4) each pixel is taken the block of 22 × 22 around F1, obtain block-based eigenmatrix F2, be i.e. the block of corresponding 3 22 × 22 of each pixel；

(5) from block-based eigenmatrix F2, training dataset and test data set are chosen:

(5a) Polarimetric SAR Image atural object being divided into 15 classes, randomly select N number of markd pixel respectively as training sample D1 from each classification, all the other markd pixels take the integer between 300～700 as test sample T1, N；

(5b) with the marginal point of Canny operator extraction Polarimetric SAR Image, training sample D1 adds the marginal point of Canny operator extraction, namely increase the training sample that confidence level is higher, training dataset D after being updated and test data set T；

(6) structure non-down sampling contourlet convolutional neural networks:

(6a) select the 8 layers of convolutional neural networks being made up of input layer → convolutional layer → pond layer → convolutional layer → pond layer → full articulamentum → full articulamentum → softmax grader, and determine the wave filter size of convolutional neural networks and the Feature Mapping figure of each layer；

(6b) replace the level 2 volume lamination in convolutional neural networks with non-down sampling contourlet transform layer, obtain non-down sampling contourlet convolutional neural networks；

(7) with non-down sampling contourlet convolutional neural networks, training dataset is trained；

(8) utilize the non-down sampling contourlet convolutional neural networks trained that test data set is classified, obtain the pixel class of each pixel in Polarimetric SAR Image test data set.

2. the Classification of Polarimetric SAR Image method of non-down sampling contourlet convolutional neural networks according to claim 1, wherein carries out denoising to Polarimetric SAR Image to be sorted in step (1), adopt exquisiteness polarization LEE filter method, and its step is as follows:

(1a) set exquisiteness polarization LEE filtering sliding window, this sliding window be sized to 5 × 5 pixels；

(1b) by sliding window in the pixel of the Polarimetric SAR Image of input, from left to right, roam from top to bottom, when often roaming a step, by sliding window according to pixel space position, from left to right, 9 subwindows it are divided into from top to bottom successively, each subwindow be sized to 3 × 3 pixels, have overlap between subwindow；

(1c) pixel value of each subwindow correspondence position is averaged, obtained average is constituted the average window of 3 × 3 pixels；

(1d) the gradient masterplate of level, vertical, the four direction of 45 degree and 135 degree is chosen, average window is weighted with four masterplates respectively, obtained weighted results is sought absolute value, selects the maximum in all absolute values, using direction corresponding for this maximum as edge direction；

(1e) from 2, the left and right subwindow of 9 subwindow Zhong Qu center window edge directions, respectively all pixel values in these 2 subwindows are averaged, the average of all pixel values of center window it is individually subtracted by 2 averages obtained, using the subwindow corresponding to value little for absolute value in average difference as direction window, wherein, center window refers to the subwindow of the 3 × 3 of 5 × 5 window center；

(1f) according to formula<1>, the weights of exquisite polarization LEE filtering are obtained:

$b=\frac{\mathrm{var}\left(y\right)-p^2{\mathrm{\&sigma;}}_v^2}{(1+{\mathrm{\&sigma;}}_v^2)\mathrm{var}\left(y\right)},---<1>$

Wherein, b represents the weights of exquisite polarization LEE filtering, and var (y) represents the variance yields of polarization SAR general power image pixel in the window of direction, and y represents the pixel of polarization SAR general power image in the window of direction, p represents the average of all pixels of polarization SAR general power image in the window of directionRepresent the variance yields of the Polarimetric SAR Image coherent speckle noise of input.

(1g) according to formula<2>, the polarization coherence matrix T of filtering after-polarization SAR image center pixel is obtained:

T=w+b (z-w),<2>

Wherein, w represents the average of the polarization coherence matrix of Polarimetric SAR Image pixel in the window of direction, and b represents the weights of exquisite LEE filtering, and z represents the polarization coherence matrix of Polarimetric SAR Image center pixel.

(1h) according to formula<3>, horizontal emission and the scattering component S of level reception can be tried to achieve_HH, Vertical Launch and vertical reception scattering component S_VV, horizontal emission and vertical reception scattering component S_HV:

$\left\{\begin{array}{c}{T}_{11}=\frac{1}{2}|S_{HH}+S_{VV}{|}^2\\ T_{22}=\frac{1}{2}|S_{HH}-S_{VV}{|}^2\\ T_{33}=2|S_{HV}{|}^2\end{array}\right.---<3>$

Wherein, T₁₁、T₂₂、T₃₃For element on the diagonal of the coherence matrix T that polarizes.

3. the Classification of Polarimetric SAR Image method of non-down sampling contourlet convolutional neural networks according to claim 1, wherein carries out Pauli decomposition to filtered polarization scattering matrix S in step (2), and its step is as follows:

(2a) define basic collision matrix, be called Pauli base: { S_a,S_b,S_c,S_d, formula is as follows:

$\left\{\begin{array}{c}{S}_a=\frac{1}{\sqrt{2}}\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right]\\ S_b=\frac{1}{\sqrt{2}}\left[\begin{array}{cc}1& 0\\ 0& -1\end{array}\right]\\ S_c=\frac{1}{\sqrt{2}}\left[\begin{array}{cc}0& 1\\ 1& 0\end{array}\right]\\ S_d=\frac{1}{\sqrt{2}}\left[\begin{array}{cc}1& -i\\ i& 1\end{array}\right]\end{array}\right.---<4>$

Wherein $\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right]$ Represent odd scattering, $\left[\begin{array}{cc}1& 0\\ 0& -1\end{array}\right]$ Represent even scattering, $\left[\begin{array}{cc}0& 1\\ 1& 0\end{array}\right]$ Represent volume scattering, $\left[\begin{array}{cc}0& -i\\ i& 0\end{array}\right]$ Representing non-existent type of ground objects, therefore d value is 0；

(2b) the Pauli base defined according to formula<4>, obtains the expression formula of polarization scattering matrix S:

$S=\left[\begin{array}{cc}{S}_{HH}& S_{HV}\\ S_{VH}& S_{VV}\end{array}\right]={\mathrm{aS}}_a+{\mathrm{bS}}_b+{\mathrm{cS}}_c+{\mathrm{dS}}_d---<5>$

The wherein value of a correspondence odd scattering, the value of b correspondence even scattering, c represents the value of volume scattering, and d represents the value of scattering composition corresponding to non-existent type of ground objects；

(2c) solve formula<5>, obtain scattering value a, b, c, d, be denoted as vector form as follows:

$K=\&lsqb;\begin{array}{cccc}a& b& c& d\end{array}\&rsqb;=\frac{1}{\sqrt{2}}{\left(\begin{array}{cccc}{S}_{HH}+S_{VV}& S_{HH}-S_{VV}& S_{HV}+S_{VH}& i\left(S_{VH}-S_{HV}\right)\end{array}\right)}^T---<6>$

When meeting reciprocity condition S_HV=S_VHTime, formula<6>is reduced to:

$K=\&lsqb;\begin{array}{ccc}a& b& c\end{array}\&rsqb;=\frac{1}{\sqrt{2}}\left(\begin{array}{ccc}{S}_{HH}+S_{VV}& S_{HH}-S_{VV}& 2S_{HV}\end{array}\right),---<7>$

The S that formula<3>is tried to achieve_HH、S_VV、S_HVSubstitution formula<7>, tries to achieve polarization characteristic K.

4. the Classification of Polarimetric SAR Image method of non-down sampling contourlet convolutional neural networks according to claim 1, wherein to the eigenmatrix F normalization based on pixel in step (3), adopt characteristic line pantography, namely first obtain the maximum max (F) of eigenmatrix F；Again by each element in eigenmatrix F all divided by maximum max (F), obtain normalized eigenmatrix F1.

5. the Classification of Polarimetric SAR Image method of non-down sampling contourlet convolutional neural networks according to claim 1, wherein the non-down sampling contourlet convolutional neural networks in step (6b), its structure is 8 layers, is expressed as:

Input layer → non-down sampling contourlet transform layer → pond layer → convolutional layer → pond layer → full articulamentum → full articulamentum → softmax grader

The parameter of each layer is:

1st layer of input layer: output characteristic mapping graph=3；

2nd layer of non-down sampling contourlet transform layer: output characteristic mapping graph=12；

3rd layer of pond layer: down-sampling yardstick=2.

4th layer of convolutional layer: output characteristic mapping graph=20, filter size=4；

5th layer of pond layer: down-sampling yardstick=2；

6th layer of full articulamentum: output characteristic mapping graph=100；

7th layer of full articulamentum: output characteristic mapping graph=64；

8th layer of softmax grader: output characteristic mapping graph=15.