CN107239751B - High-resolution SAR image classification method based on non-subsampled contourlet full convolution network - Google Patents

Abstract

A high-resolution SAR image classification method based on a non-subsampling contourlet full convolution network, comprising inputting a high-resolution SAR image to be classified, performing multi-layer non-subsampling contourlet transformation on each pixel in the image, and obtaining each pixel The low-frequency coefficient and high-frequency coefficient of the point; the low-frequency coefficient and the high-frequency coefficient are selected and fused to form a pixel-based feature matrix F; the element values in the feature matrix F are normalized to obtain the normalized feature matrix F1; The normalized feature matrix F1 is cut into blocks, and the feature block matrix F2 is obtained as sample data; the training data set feature matrix W1 and the test data set feature matrix W2 are constructed; the classification model based on the fully convolutional neural network is constructed; The trained model classifies the test data set T, obtains the category of each pixel in the test data set T, and compares the obtained category of each pixel with the class chart to calculate the classification accuracy, which improves the classification accuracy and speed. .

Description

High-resolution SAR image classification method based on non-subsampled contourlet fully convolutional network

technical field

The invention belongs to the field of image processing, and in particular relates to a high-resolution SAR image classification method based on a non-subsampling contourlet full convolution network, which can be applied to high-resolution SAR images and effectively improves the recognition accuracy of targets.

Background technique

Synthetic Aperture Radar (SAR) is a kind of remote sensing sensor that has been widely studied and applied in recent years. Compared with other sensors such as optical and infrared, SAR imaging is not limited by conditions such as weather and illumination, and can be used for target for all-weather, all-day reconnaissance. Moreover, SAR also has a certain penetration ability, which can detect the target under unfavorable conditions such as cloud interference, tree occlusion, or shallow buried surface of the target. In addition, due to the special imaging mechanism of SAR, high-resolution SAR images contain different content from other sensors, providing more comprehensive information for target detection. Because SAR has many remarkable advantages, it has great application potential. In recent years, the research on SAR technology has attracted extensive attention, and many research results have been successfully applied to environmental monitoring, terrain measurement, target detection and so on.

The key to high-resolution SAR image classification is the target feature extraction of high-resolution SAR images. Existing SAR image classification techniques include statistical-based classification methods, image texture-based classification methods, and deep learning-based classification methods.

The classification method based on statistics is to classify according to the difference of statistical characteristics of image regions of different properties, but this method ignores the spatial distribution characteristics of the image, so the classification results are often unsatisfactory. In recent years, some classification methods based on texture features have also appeared, such as the method based on gray level co-occurrence matrix (GLCM), the method based on Markov random field (MRF), the Gabor wavelet method, etc. However, due to the mechanism of coherent imaging of SAR images, resulting in The texture in SAR images is not obvious and not robust. In addition, the computer texture features need to scan the image point by point, which requires a huge amount of calculation and cannot meet the real-time requirements.

The above traditional SAR image classification methods can only rely on manual extraction of some shallow features representing target characteristics. These shallow features are only obtained by converting the original input signal to a specific problem space, and cannot fully characterize the difference between target pixels. neighborhood correlation. In 2006, Hinton et al. proposed an unsupervised layer-by-layer greedy training method to solve the "gradient dissipation" problem caused by increasing depth. Subsequently, many scholars proposed a variety of DL models according to different application backgrounds, such as Deep Belief Network (DBN), Stacked Denoising Autoencoders (SDA) and Convolutional Neural Network (Convolutional Neural Network) , CNN) etc. However, none of the above feature extraction methods take into account the multi-scale, multi-directional, and multi-resolution characteristics of high-resolution SAR images. Therefore, it is difficult to obtain high classification accuracy for high-resolution SAR images with complex backgrounds.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a high-resolution SAR image classification method based on a non-subsampling contourlet full convolution network in view of the above-mentioned problems in the prior art, combined with the multi-scale, multi-directional, and multi-resolution characteristics of the high-resolution SAR image , improve the accuracy and classification speed of its image classification, and then effectively improve the recognition accuracy of the target.

In order to achieve the above object, the technical solution adopted in the present invention comprises the following steps:

1) Input the high-resolution SAR image to be classified, perform multi-layer non-subsampling contourlet transformation on each pixel in the image, and obtain the low-frequency and high-frequency coefficients of each pixel;

2) Select and fuse low-frequency coefficients and high-frequency coefficients to form a pixel-based feature matrix F;

3) Normalize the element values in the feature matrix F to between [0, 1] to obtain a normalized feature matrix F1;

4) dicing the normalized feature matrix F1 into blocks to obtain a feature block matrix F2 and use it as sample data;

5) Construct training data set feature matrix W1 by training data set D, and construct test data set feature matrix W2 by testing data set T;

6) Construct a classification model based on a fully convolutional neural network;

7) The classification model is trained through the training data set D to obtain a trained model;

8) Use the trained model to classify the test data set T, obtain the category of each pixel point in the test data set T, and compare the obtained category of each pixel point with the class map to calculate the classification accuracy.

Described step 1) carries out three-layer non-subsampling contourlet transformation to each pixel in the image; non-subsampling contourlet transformation includes non-subsampling pyramid decomposition and non-subsampling directional filter decomposition, and the non-subsampling Pyramid decomposition decomposes the time-frequency plane into a low-frequency sub-generation and multiple annular high-frequency sub-generations through a non-subsampling filter bank. The coefficients of the subimage.

Described step 2) sort the high-frequency coefficients from large to small, select the first 50% of the high-frequency coefficients, and fuse them with the low-frequency coefficients after the third layer transformation, and define the pixel-based feature matrix F size is M1 ×M2×1, M1 is the length of the SAR image to be classified, M2 is the width of the SAR image to be classified, and the fusion result is assigned to the pixel-based feature matrix F.

The normalization described in step 3) is realized by a feature linear scaling method, a feature standardization method or a feature whitening method; the feature linear scaling method first obtains the maximum value max(F) of the pixel-based feature matrix F; Each element in the feature matrix F of the point is divided by the maximum value max(F) to obtain the normalized feature matrix F1.

In the step 4), the normalized feature matrix F1 is cut into blocks according to the size of 128×128 and the interval of 50.

The concrete operation of described step 5) is as follows:

5a) Divide the SAR image features into three categories, record the positions of the pixels corresponding to each category in the image to be classified, and generate three positions A1, A2, A3;

5b) Randomly select 5% of the elements from A1, A2, and A3 to generate three pixel positions B1, B2, and B3 that correspond to different types of ground objects and are selected as the training data set, where B1 corresponds to the first type The position of the pixels in the ground objects selected as the training data set in the image to be classified, B2 is the position of the pixels in the image to be classified corresponding to the second type of ground objects selected as the training data set, and B3 is the corresponding 3 types of features are selected as the positions of the pixels in the training data set in the image to be classified, and the elements in B1, B2, and B3 are combined to form the position L1 of all the pixels in the training data set in the image to be classified;

5c) Use the remaining 95% of the elements in A1, A2, and A3 to generate 3 pixel positions C1, C2, and C3 corresponding to different types of ground objects that are selected as the test data set, where C1 is the ground object corresponding to the first type. The position of the pixel selected as the test data set in the image to be classified, C2 is the position of the pixel selected as the test data set in the image to be classified corresponding to the second type of terrain, and C3 is the corresponding to the third type of ground. is selected as the position of the pixels in the test data set in the image to be classified, and the elements in C1, C2, and C3 are combined to form the position L2 of all the pixels in the test data set in the image to be classified;

5d) define the training data set feature matrix W1 of the training data set D, take the value on the corresponding position according to L1 in the feature block matrix F2, and assign it to the training data set feature matrix W1 of the training data set D;

5e) Define the test data set feature matrix W2 of the test data set T, take the value at the corresponding position in the feature block matrix F2 according to L2, and assign it to the test data set feature matrix W2 of the test data set T.

Described step 6) constructing the classification model based on full convolutional neural network comprises the following steps:

6a) Choose a sequence consisting of input layer, convolution layer, pooling layer, convolution layer, pooling layer, convolution layer, pooling layer, convolution layer, pooling layer, convolution layer, dropout layer, convolution layer Layer, Dropout layer, convolution layer, deconvolution upsampling layer, Crop cropping layer, softmax classifier composed of 17 layers of deep neural network, the parameters of each layer are as follows:

For the first input layer, set the number of feature maps to 3;

For the second convolutional layer, set the number of feature maps to 32 and the convolution kernel size to 5×5;

For the third layer pooling layer, set the downsampling size to 2;

For the fourth convolutional layer, set the number of feature maps to 64 and the convolution kernel size to 5×5;

For the 5th layer pooling layer, set the downsampling size to 2;

For the sixth convolutional layer, set the number of feature maps to 96 and the convolution kernel size to 3×3;

For the 7th layer pooling layer, set the downsampling size to 2;

For the 8th convolutional layer, set the number of feature maps to 128 and the convolution kernel size to 3×3;

For the 9th pooling layer, set the downsampling size to 2;

For the 10th convolutional layer, set the number of feature maps to 128 and the convolution kernel size to 3×3;

For the 11th Dropout layer, set the sparsity coefficient to 0.5;

For the 12th convolutional layer, set the number of feature maps to 128 and the convolution kernel size to 1×1;

For the 13th Dropout layer, set the sparsity coefficient to 0.5;

For the 14th convolutional layer, set the number of feature maps to 2 and the convolution kernel size to 1×1;

For the 15th deconvolution upsampling layer, set the number of feature maps to 2 and the convolution kernel size to 32×32;

For the 16th Crop layer, set the final crop size to 128×128;

For the 17th layer Softmax classifier, set the number of feature maps to 2;

6b) Set the size of the convolution kernel of the second convolutional layer to 5×5 to reduce the receptive field.

Described step 7) take the training data set feature matrix W1 as the input of the classification model, take the category of each pixel in the training data set D as the output of the classification model, and solve the error between the above-mentioned categories and the correct categories marked manually , and back-propagates the error, optimizes the network parameters of the classification model, and obtains a trained classification model.

The step 8) takes the test data set feature matrix W2 as the input of the trained classification model, and the output result of the trained classification model is the classification category obtained by classifying each pixel point in the test data set T.

Compared with the prior art, the present invention has the following beneficial effects: by extending image block features into pixel-level features, repeated storage and calculation convolution caused by using pixel blocks are avoided, and the speed and efficiency of classification are improved. . Due to the introduction of multi-layer non-subsampling contourlet transform before the fully convolutional neural network, low-frequency coefficients and high-frequency coefficients are obtained. Therefore, the low-frequency coefficients are more capable of classifying and discriminating than the high-frequency coefficients. The invention selects and fuses low-frequency coefficients and high-frequency coefficients to improve the classification accuracy. Since the fully connected layer in the convolutional neural network is replaced by a deconvolutional layer, it can accept input images of any size without requiring all Both the training and test images have the same size. To sum up, the high-resolution SAR image classification method of the present invention can not only improve the classification accuracy, but also improve the classification speed.

Description of drawings

Fig. 1 is the flow chart of the classification method of the present invention;

Fig. 2 is the manual labeling diagram of the image to be classified according to the present invention;

FIG. 3 is a diagram of the classification result of the image to be classified according to the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings.

Referring to Fig. 1, the implementation steps of the image classification method of the present invention are as follows:

Step 1. Input the high-resolution SAR image to be classified, and perform 3-layer non-subsampling contourlet transform on each pixel to obtain its high and low frequency coefficients; the high-resolution SAR image to be classified is the German Space Agency (DLR) F-SAR The X-band horizontal polarization map obtained by the aviation system in 2007 has a resolution of 1m and an image size of 6187*4278.

1a) Transform the classification features of each pixel to obtain transform coefficients, and the transform methods include wavelet transform, non-subsampling stationary wavelet transform, curvelet transform, non-subsampling contourlet transform and other methods;

1b) This example uses non-subsampling contourlet transform to perform 3-layer transformation on each pixel, and the non-subsampling contourlet transform includes non-subsampling pyramid (NSP) decomposition and non-subsampling directional filter (NSDFB);

1c) The non-subsampling pyramid (NSP) transform utilizes non-subsampling filter banks (NSFs) to decompose the time-frequency plane into a low-frequency descendant and many annular high-frequency descendants;

1d) The non-subsampling directional filter (NSDFB) is a two-channel non-subsampling filter bank;

In this example, the image is filtered by 3-level NSP, and the coefficients of 1 low-pass image and 3 band-pass images are obtained;

After the image in this example is multi-scale decomposed by NSP, the band-pass image of the image is further decomposed by NSDFB at levels 0, 1, and 3 in multiple directions, thereby obtaining 1, 2, and 8 bandpass sub-image coefficients respectively.

Step 2, select and fuse high and low frequency coefficients to form a pixel-based feature matrix F. In this example, the high-frequency coefficients obtained by decomposing are sorted from large to small, and the top 50% of them are selected and fused with the low-frequency coefficients after the third layer transformation as the transform domain classification feature. A matrix of size M1×M2×1 is defined, and the fusion result is assigned to the matrix to obtain the pixel-based feature F, where M1 is the length of the SAR image to be classified, and M2 is the width of the SAR image to be classified.

Step 3, normalize the pixel-based feature matrix F.

Commonly used normalization methods are: feature linear scaling method, feature standardization and feature whitening.

In this example, the feature linear scaling method is used, that is, the maximum value max(F) of the pixel-based feature matrix F is firstly obtained; then each element in the pixel-based feature matrix F is divided by the maximum value max(F) , get the normalized feature matrix F1.

In step 4, the normalized feature matrix F1 is cut into blocks according to the size of 128×128 and the interval of 50 to form a small feature block matrix F2 as sample data.

Step 5: Construct training data set feature matrix W1 by training data set D, and construct test data set feature matrix W2 by testing data set T; specifically include the following steps:

5a) Divide the SAR image features into 3 categories, record the positions of the pixels corresponding to each category in the image to be classified, and generate 3 kinds of positions A1, A2, A3 corresponding to different categories of pixels;

Among them, A1 corresponds to the position of the first type of ground object pixels in the image to be classified, A2 corresponds to the position of the second type of ground object pixels in the to-be-classified image, and A3 corresponds to the third type of ground object pixels in the to-be-classified image. s position;

5b) randomly select 5% of the elements from the pixel positions A1, A2, and A3 of the different types of ground objects to generate three positions B1, B2, and B3 corresponding to the different types of ground objects selected as the pixel points of the training data set;

Among them, B1 is the position in the image to be classified corresponding to the first type of ground object selected as the training data set, and B2 is the pixel point corresponding to the second type of ground object selected as the training data set in the to-be-classified image. B3 is the position of the pixel in the image to be classified corresponding to the third type of ground object selected as the training data set, and the elements in B1, B2, and B3 are combined to form all the pixels in the training data set. position L1 in the image to be classified;

5c) Use the remaining 95% of the elements in A1, A2, and A3 to generate 3 types of locations C1, C2, and C3 of the pixels corresponding to different types of ground objects selected as the test data set, where C1 is the ground object corresponding to the first type The position of the pixels selected as the test data set in the image to be classified, C2 is the position of the pixels selected as the test data set in the image to be classified corresponding to the second category of features, and C3 is the corresponding to the third category. The position of the pixels in the ground object selected as the test data set in the image to be classified, and the elements in C1, C2, and C3 are combined to form the position L2 of all the pixels in the test data set in the image to be classified;

5d) define the training data set feature matrix W1 of the training data set D, take the value on the corresponding position according to L1 in the feature matrix F2 based on the image block, and assign it to the training data set feature matrix W1;

5e) Define the test data set feature matrix W2 of the test data set T, take the value at the corresponding position in the feature block matrix F2 according to L2, and assign it to the test data set feature matrix W2.

Step 6, construct a classification model based on a fully convolutional neural network.

6a) Select a layer consisting of input layer→convolutional layer→pooling layer→convolutional layer→pooling layer→convolutional layer→pooling layer→convolutional layer→pooling layer→convolutional layer→Dropout layer→convolutional layer →Dropout layer→Convolutional layer→Upsampling layer (deconvolution)→Crop layer (cropping)→17-layer deep neural network composed of softmax classifier, the parameters of each layer are as follows:

For the first input layer, set the number of feature maps to 3;

For the second convolutional layer, set the number of feature maps to 32 and the convolution kernel size to 5×5;

For the third layer pooling layer, set the downsampling size to 2;

For the fourth convolutional layer, set the number of feature maps to 64 and the convolution kernel size to 5×5;

For the 5th layer pooling layer, set the downsampling size to 2;

For the sixth convolutional layer, set the number of feature maps to 96 and the convolution kernel size to 3×3;

For the 7th layer pooling layer, set the downsampling size to 2;

For the 8th convolutional layer, set the number of feature maps to 128 and the convolution kernel size to 3×3;

For the 9th pooling layer, set the downsampling size to 2;

For the 10th convolutional layer, set the number of feature maps to 128 and the convolution kernel size to 3×3;

For the 11th Dropout layer, set the sparsity coefficient to 0.5;

For the 12th convolutional layer, set the number of feature maps to 128 and the convolution kernel size to 1×1;

For the 13th Dropout layer, set the sparsity coefficient to 0.5;

For the 14th convolutional layer, set the number of feature maps to 2 and the convolution kernel size to 1×1;

For the 15th layer upsampling layer, set the number of feature maps to 2 and the convolution kernel size to 32×32;

For the 16th Crop layer, set the final crop size to 128×128;

For the 17th layer Softmax classifier, set the number of feature maps to 2.

6b) Set the size of the convolution kernel of the second convolutional layer to 5×5 to reduce the receptive field;

Step 7: Train the classification model with the training data set to obtain a trained classification model.

The training dataset feature matrix W1 is used as the input of the classification model, and the category of each pixel in the training dataset D is used as the output of the classification model. Propagation, optimize the network parameters of the classification model, and obtain the trained classification model. The correct class label of manual labeling is shown in Figure 2.

Step 8: Use the trained classification model to classify the test data set.

The test data set feature matrix W2 of the test data set T is used as the input of the trained classification model, and the output of the trained classification model is the classification category obtained by classifying each pixel point in the test data set.

The effect of the present invention is further illustrated by the following simulation experiments:

1. Simulation conditions:

The hardware platform is: HPZ840.

The software platform is: Caffe.

2. Simulation content and results:

Experiments are carried out under the above simulation conditions with the method of the present invention, that is, 5% of the marked pixels are randomly selected from the SAR data as training samples, and the remaining marked pixels are used as test samples, and the classification result as shown in Figure 3 is obtained.

It can be seen from Figure 3 that the regional consistency of the classification results is good, and the edges of the three categories of farmland, forest and town are also clearer, and the detailed information is maintained.

Then reduce the training samples in turn, so that the training samples account for 4%, 3%, and 2% of the total number of samples, and compare the classification accuracy of the present invention and the test data set of the full convolutional neural network. The results are shown in Table 1:

Table 1

The proportion of training samples FCN-8 this invention 5% 94.0039% 94.3360% 4% 93.3933% 94.1524% 3% 92.6727% 93.3117% 2% 91.4413% 92.4162%

It can be seen from Table 1 that when the training samples account for 5%, 4%, 3%, and 2% of the total number of samples, the classification accuracy of the test data set of the present invention is higher than that of the pure full convolutional neural network. In summary, the present invention, by introducing non-subsampling contourlet transform into the fully convolutional neural network, takes into account the directional information and spatial information of the high-resolution SAR image, effectively improves the expression ability of image features, and enhances the generalization ability of the model , so that high classification accuracy can still be achieved in the case of fewer training samples.

Claims (8)

1. A high-resolution SAR image classification method based on a non-downsampling contourlet full convolution network is characterized by comprising the following steps:

1) inputting a high-resolution SAR image to be classified, and carrying out multilayer non-downsampling contourlet transformation on each pixel point in the image to obtain a low-frequency coefficient and a high-frequency coefficient of each pixel point;

2) selecting and fusing the low-frequency coefficient and the high-frequency coefficient to form a characteristic matrix F based on pixel points;

3) normalizing the element values in the feature matrix F to be between [0 and 1] to obtain a normalized feature matrix F1;

4) dicing the normalized feature matrix F1 to obtain a feature block matrix F2 as sample data;

5) constructing a training data set characteristic matrix W1 through a training data set D, and constructing a test data set characteristic matrix W2 through a test data set T; the specific operation is as follows:

5a) dividing the high-resolution SAR image surface features into 3 classes, recording the positions of pixel points corresponding to each class in the image to be classified, and generating three positions A1, A2 and A3 respectively representing the positions of the pixel points of the three classes of surface features in the image to be classified;

5b) randomly selecting 5% of elements from the A1, A2 and A3 to generate three pixel positions B1, B2 and B3 which correspond to different types of ground objects and are selected as training data sets, wherein B1 corresponds to the position of a pixel selected as a training data set in a1 st type of ground object in an image to be classified, B2 corresponds to the position of a pixel selected as a training data set in a2 nd type of ground object in the image to be classified, B3 corresponds to the position of a pixel selected as a training data set in A3 rd type of ground object in the image to be classified, and combining the elements from the B1, the B2 and the B3 to form a position L1 of all pixels of the training data set in the image to be classified;

5c) generating 3 pixel point positions C1, C2 and C3 corresponding to different types of ground objects and selected as test data sets by using the rest 95% of the elements in the A1, A2 and A3, wherein C1 is the position of a pixel point selected as a test data set in a1 st type of ground object in an image to be classified, C2 is the position of a pixel point selected as a test data set in a2 nd type of ground object in the image to be classified, C3 is the position of a pixel point selected as a test data set in A3 rd type of ground object in the image to be classified, and combining the elements in the C1, C2 and C3 to form the position L2 of all the pixel points of the test data set in the image to be classified;

5d) defining a training data set characteristic matrix W1 of a training data set D, taking values at corresponding positions in a characteristic block matrix F2 according to L1, and assigning the values to a training data set characteristic matrix W1 of the training data set D;

5e) defining a test data set characteristic matrix W2 of the test data set T, taking values at corresponding positions in a characteristic block matrix F2 according to L2, and assigning the values to a test data set characteristic matrix W2 of the test data set T;

6) constructing a classification model based on a full convolution neural network;

7) training the classification model through a training data set D to obtain a trained model;

8) classifying the test data set T by using the trained model to obtain the category of each pixel point in the test data set T, comparing the obtained category of each pixel point with the category label chart, and calculating the classification accuracy.

2. The method for classifying the high-resolution SAR image based on the non-subsampled contourlet full convolution network according to claim 1, wherein the method comprises the following steps: step 1) carrying out three-layer non-downsampling contourlet transformation on each pixel point in an image; the non-downsampling contourlet transformation comprises non-downsampling pyramid decomposition and non-downsampling direction filter decomposition, the non-downsampling pyramid decomposition decomposes a time-frequency plane into a low-frequency filial generation and a plurality of annular high-frequency filial generations through a non-downsampling filter bank, and a band-pass image formed by the non-downsampling pyramid decomposition is decomposed through the non-downsampling direction filter to obtain coefficients of the band-pass sub-images.

3. The method for classifying the high-resolution SAR image based on the non-subsampled contourlet full convolution network according to claim 2, wherein the method comprises the following steps: and 2) sequencing the high-frequency coefficients from large to small, selecting the high-frequency coefficients of the first 50% of the high-frequency coefficients, fusing the high-frequency coefficients with the low-frequency coefficients after the third layer of transformation, defining the size of a characteristic matrix F based on the pixel points as M1 multiplied by M2 multiplied by 1, defining the size of the characteristic matrix F based on the pixel points as M1 as the length of the SAR image to be classified and defining the size of the characteristic matrix F based on the pixel points as M2 as the width of the SAR image to be classified, and assigning the fusion.

4. The method for classifying the high-resolution SAR image based on the non-subsampled contourlet full convolution network according to claim 1, wherein the method comprises the following steps: the normalization in the step 3) is realized by a characteristic linear scaling method, a characteristic standardization method or a characteristic whitening method; firstly, solving the maximum value max (F) of a characteristic matrix F based on pixel points by a characteristic linear scaling method; and dividing each element in the characteristic matrix F based on the pixel points by the maximum value max (F) to obtain a normalized characteristic matrix F1.

5. The method for classifying the high-resolution SAR image based on the non-subsampled contourlet full convolution network according to claim 1, wherein the method comprises the following steps: step 4) the normalized feature matrix F1 is diced at intervals of 50 and size 128 x 128.

6. The method for classifying the high-resolution SAR image based on the non-subsampled contourlet full convolution network as claimed in claim 1, wherein the step 6) of constructing the classification model based on the full convolution neural network comprises the following steps:

6a) selecting a 17-layer deep neural network consisting of an input layer, a convolutional layer, a pooling layer, a convolutional layer, a Dropout layer, a convolutional layer, a deconvolution up-sampling layer, a Crop layer and a softmax classifier in sequence, wherein the parameters of each layer are as follows:

setting the number of feature maps to be 3 for the input layer of the layer 1;

for the 2 nd convolutional layer, setting the number of feature maps to be 32 and the size of a convolutional kernel to be 5 multiplied by 5;

for the 3 rd pooling layer, the down-sampling size is set to 2;

for the 4 th convolutional layer, setting the number of feature maps to be 64 and the size of a convolutional kernel to be 5 multiplied by 5;

for the 5 th pooling layer, the down-sampling size is set to 2;

for the convolution layer of the layer 6, setting the number of feature maps to be 96 and the size of a convolution kernel to be 3 multiplied by 3;

for the 7 th pooling layer, the downsampling size is set to 2;

for the 8 th convolutional layer, setting the number of feature maps to be 128 and the size of a convolutional kernel to be 3 multiplied by 3;

setting the downsampling size to be 2 for the 9 th pooling layer;

for the 10 th convolutional layer, setting the number of feature maps to be 128 and the size of a convolutional kernel to be 3 multiplied by 3;

setting the sparsity factor to 0.5 for the 11 th Dropout layer;

for the 12 th convolutional layer, setting the number of feature maps to be 128 and the size of a convolutional kernel to be 1 multiplied by 1;

for the 13 th Dropout layer, setting the sparsity factor to 0.5;

for the 14 th convolutional layer, setting the number of feature maps to be 2 and the size of a convolutional kernel to be 1 multiplied by 1;

for the 15 th deconvolution up-sampling layer, setting the number of feature maps to be 2 and the convolution kernel size to be 32 multiplied by 32;

setting the final cutting specification to be 128 multiplied by 128 for the 16 th Crop layer;

setting the number of feature maps to be 2 for the 17 th layer Softmax classifier;

6b) the convolution kernel size of the second layer of convolutional layers is set to 5 x 5, reducing the receptive field.

7. The method for classifying the high-resolution SAR image based on the non-subsampled contourlet full convolution network as claimed in claim 1, wherein the step 7) is to use a training data set feature matrix W1 as an input of a classification model, use the category of each pixel point in a training data set D as an output of the classification model, solve an error between the category and a correct category of an artificial marker, perform back propagation on the error, and optimize the network parameters of the classification model to obtain the trained classification model.

8. The method for classifying the high-resolution SAR image based on the non-subsampled contourlet full convolution network as claimed in claim 7, wherein the step 8) is to input a test data set feature matrix W2 as a trained classification model, and an output result of the trained classification model is a classification category obtained by classifying each pixel point in the test data set T.

Images