Polarization SAR image classification based on CNN and SVM

An image and classifier technology, applied in the field of image processing, can solve the problems of small number of features, underutilization of polarization information, arbitrary division of regions, etc., and achieve the effect of improving classification accuracy

Active Publication Date: 2015-12-23
XIDIAN UNIV
2 Cites 45 Cited by

AI-Extracted Technical Summary

Problems solved by technology

[0005] Unsupervised classification methods include: H/α unsupervised classification proposed by Cloude et al., which obtains scattering entropy H and average scattering angle α characteristic parameters through Cloude target decomposition, and performs eight classifications on targets according to the range of these two parameters. In this method, the classification boundary is fixed and the division of regions is too arbitrary, and only the two parameters H and α are used, and the polarization information is not fully utilized, resulting in low classification accuracy; Lee et al. proposed the H method based on Cloude target decomposition and Wishart classifier. /α-Wishart unsupervised classification method, which adds Wishart iteration on the basis of the original H/α classification, which makes up for the defects of the fixed boundary of H/α classification, but this method cannot well maintain all kinds of polarization scattering Characteristics; Lee et al. proposed a polarimetric SAR...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Method used

6b) the AE hidden layer feature obtained by 6a) is sent into the softmax classifier, because the principle of the softmax classifier is simple, it is easy to realize the fine-tuning of the AE network parameters, according to the difference between the classification result and the reference map label, using the reverse Fine-tune the parameters of the AE network to the propagation algorithm, so that the parameters of the AE network can be further optimized, so as to obtain better parameters of the CNN convolution layer.
Among the present invention, step 4,5,6 is the study to CNN convolutional layer parameter, and it obtains CNN convolutional layer parameter based on the study of AE network, then fine-tunes based on softmax classifier in supervised mode again, avoids The traditional CNN convolutional layer parameter learning takes a long time and is easy to fall into the defect of local optimum.
Step 1, the polarization SAR image to be classified after input filtering, referring to Fig. 2, obtains polarization coherence matrix T, and this polarization SAR image is a pair of Dutch farmland figure, mainly comprises rapeseed, sugar beet, potato, Alfalfa, grassland, wheat, peas and other crops and a piece of bare land, different colors in the figure represent different categories, the present invention is aimed at classifying this polarimetric SAR image, and the polarimetric SAR image to be classified itself has ground objects For the distribution reference map, see Fig. 3. In the figure, some pixels are marked with categories; the image filtering to be classified is mainly performed by polarization refined Lee filtering, which effectively removes the coherent speckle noise of the polarimetric SAR image.
Step 8, with all the pixel points marked in the ground object distribution reference figure of polarimetric SAR image as benchmark, according to the sampling rate of no more than 10%, randomly select the pixel point of every class, as training sample set, corresponding Compared with the fixed selection of image blocks of each type as the traini...
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Abstract

The invention discloses a polarization SAR image classification method based on CNN and SVM, and mainly aims to solve the problem of the existing polarization SAR image classification method that the classification precision is low. The method comprises the steps as follows: (1) inputting a to-be-classified polarization SAR image after filtering; extracting and normalizing the original feature of each pixel point based on a polarization coherence matrix and by taking the neighborhood into consideration; training an AE network, and obtaining the parameter of a CNN convolution layer through softmax fine-tuning; setting a CNN pooling layer as average pooling, and determining the parameter of the CNN pooling layer; and sending the features of CNN learning to an SVM for classification to obtain the classification result of the polarization SAR image. Compared with the existing methods, the spatial correlation of the image is fully considered, a new neighborhood processing method is proposed based on CNN, features more conductive to polarization SAR image classification can be extracted, the classification accuracy is obviously improved, and the method can be used for polarization SAR image surface feature classification and object identification.

Application Domain

Technology Topic

Image

  • Polarization SAR image classification based on CNN and SVM
  • Polarization SAR image classification based on CNN and SVM
  • Polarization SAR image classification based on CNN and SVM

Examples

  • Experimental program(8)

Example Embodiment

[0035] Example 1
[0036] The present invention is a polarization SAR image classification method based on CNN and SVM, see figure 1 , Polarization SAR image classification includes the following steps:
[0037] Step 1. Input the filtered polarized SAR image to be classified, see figure 2 , To obtain the polarization coherence matrix T, the polarization SAR image is a Dutch farmland map, mainly including rapeseed, sugar beet, potato, alfalfa, grassland, wheat, peas and other crops and a piece of bare land. Different colors in the picture represent different The category of this example is to classify this polarized SAR image, and the polarized SAR image to be classified is attached with a reference map for the distribution of ground features, see image 3 In the figure, some pixels are classified; the image filtering to be classified is mainly the polarization refined Lee filtering.
[0038] Step 2: Based on the polarization coherence matrix T of the polarization SAR image, obtain the original feature of each pixel of the image, and normalize it to [0.1, 0.9], so that different features in the original feature have the same dimension, which is convenient for subsequent follow-up operating. In this step, the present invention is based on the polarization coherence matrix, and the neighborhood information is taken into consideration, and the original feature of each pixel is extracted and normalized.
[0039] Step 3: Randomly select 10% of the pixels in the polarization SAR image to be classified as the training sample; in order to speed up, the present invention selects only 10% of the pixels as the training sample.
[0040] Step 4. Use the training samples obtained in step 3 to train the AE network to obtain AE network parameters; CNN includes multiple convolutional layers and multiple pooling layers. The CNN structure used in the present invention only includes one convolutional layer and one The layer pooling layer, because no matter what distribution the training samples obey, the AE network can quickly and conveniently express it well. In the present invention, the CNN convolutional layer parameters are determined by training the AE network. The present invention proposes a new neighborhood processing method based on CNN to perform convolution operations on image blocks of training samples.
[0041] Step 5: Train the AE network repeatedly based on the AE network parameters obtained in Step 4 until the termination condition is met. After the termination condition is satisfied, the parameters of the AE network obtained are used as the initial convolutional neural network (CNN) convolutional layer parameter θ =(W,b), where W is the convolution kernel of the CNN convolution layer. In this example, the size of the convolution kernel is set to 5×5, the number is 20, and b is the bias of the CNN convolution layer; the present invention The termination condition selected in is that the objective function value of the AE network changes less than 10 -9.
[0042] Step 6. Use the softmax classifier and the ground feature distribution of the polarized SAR image to be classified with reference to the marked pixels in the image, fine-tune the initial CNN convolutional layer parameters, and use the fine-tuned parameters as the trained CNN convolution Layer parameters
[0043] Steps 4, 5, and 6 of the present invention are the learning of the CNN convolutional layer parameters, where the local perception area of ​​the CNN convolutional layer takes into account the spatial correlation of the image, and the CNN convolutional layer parameters are obtained based on the learning of the AE network. Based on the softmax classifier, the CNN convolutional layer parameters are fine-tuned in a supervised manner, avoiding the traditional CNN convolutional layer parameter learning that takes a long time and is easy to fall into the local optimum.
[0044] Step 7. The CNN structure of the present invention includes a convolutional layer and a pooling layer: the pooling mode of the CNN pooling layer is set to average pooling. In this example, the pooling size is 2×2, and the pooling size is The determination of is related to the image size. As the image size increases, the pooling size can be appropriately expanded.
[0045] Steps 3, 4, 5, 6, and 7 are the learning of the CNN structure of the present invention. Step 7 is to determine the parameters of the CNN pooling layer. The CNN pooling methods mainly include average pooling and maximum pooling. Because of average pooling It has stable characteristics, so the average pooling is adopted in the present invention.
[0046] Step 8. Taking the ground feature distribution of the polarized SAR image as a reference to all the marked pixels in the reference map, randomly select each type of pixel as a training sample set according to a sampling rate of no more than 10%, compared to the fixed Select each type of image block as the training sample set. Random selection is more flexible, the selected pixels are more representative, and more robust to noise, which improves the generalization ability; the pixels in the training sample set All have training sample labels. The selected training sample set is to train the SVM classifier. The present invention randomly selects the pixels of each class according to the sampling rate of no more than 10%, which not only speeds up the training speed, but also shows that in the case of small samples, the SVM classifier The generalization ability is already very good.
[0047] Step 9. Send the original features of each pixel in the training sample set to the trained CNN structure to obtain the corresponding CNN features of each pixel; sending the trained CNN structure includes sending the trained CNN volume first Build up the layers to get the CNN convolution map, and then send the CNN convolution map to the trained CNN pooling layer.
[0048] Step 10. Use the CNN features and training sample labels of the training sample set obtained in step 9 to train a support vector machine (SVM) classifier, and then use the trained SVM classifier to classify all remaining pixels of the polarized SAR image one by one : Take all the remaining pixels of the polarized SAR image to be classified as the test sample, and send the original feature of each pixel in the test sample into the trained CNN structure to obtain the CNN feature of each pixel in the test sample, and then These features are input to the trained SVM for classification, and the category label of each pixel is obtained; Steps 8, 9, and 10 are the learning of the SVM classifier in the present invention.
[0049] Step 11. Output the classified polarization SAR image, see Image 6 , Image 6 The gray area in the upper left corner of the middle is rapeseed, the black area in the upper left is bare ground, the dark area in the lower middle represents beets, and the dark gray area in the middle is potatoes. figure 2 , Image 6 The areas of rapeseed, bare land, sugar beet and potatoes are largely related to figure 2 Remaining consistent, the present invention can not only correctly classify these regions with few stray points and no serious misclassification phenomenon, but also has good region retention and calculates the classification accuracy.
[0050] The present invention adopts a new feature learning method—CNN. Its convolutional layer performs convolution operations on image blocks of input data, and its pooling layer performs averaging operations on the convolutional images obtained by the convolutional layer, which can prevent excessive Fitting, in this way, the original features of the polarized SAR image can be deep-learned again, and the features that are more conducive to the classification of the polarized SAR image can be extracted.

Example Embodiment

[0051] Example 2
[0052] The polarization SAR image classification method based on CNN and SVM is the same as that in Example 1. In step 2, the original feature of each pixel of the image is obtained according to the following steps:
[0053] 2a) Since the polarization coherence matrix T has rich phase and amplitude information about the radar target, and both are positive semi-definite Hamiltonian matrices, the diagonal elements of the polarization coherence matrix T with a dimension of 3×3 can be extracted, And the real and imaginary parts of the 3 elements except the diagonal elements in the upper triangle position. Each pixel has a total of 9-dimensional features, expressed as a 3×3 matrix
[0054] T 11 T twenty two T 33 r e a l ( T 12 ) i m a g ( T 12 ) r e a l ( T 13 ) i m a g ( T 13 ) r e a l ( T twenty three ) i m a g ( T twenty three )
[0055] Where T 11 ,T 22 ,T 33 Is the diagonal element of the polarization coherence matrix T, real(T 12 ),real(T 13 ),real(T 23 ),imag(T 12 ),imag(T 13 ),imag(T 23 ) Are the real and imaginary parts of the three elements at the upper triangular position of the polarization coherence matrix T;
[0056] 2b) Since the present invention is based on CNN, it is required to input in the image format, that is, in the form of a matrix, and each pixel of the polarized SAR image is a sample, and considering the spatial correlation of the image, the present invention aims at each For an element, expand the N×N neighborhood to get (3N) 2 The dimensional feature is expressed in the form of a 3N×3N matrix, and the last column and the last row are filled with mirror symmetry. The original feature of each pixel is expressed in the form of a (3N+1)×(3N+1) matrix. In this example, the 5×5 neighborhood is expanded to obtain 225-dimensional original features, which are represented in the form of a 15×15 matrix. The last column and the last row are filled with mirror symmetry. The original feature of each pixel is 16×16 Represented in matrix form.

Example Embodiment

[0057] Example 3
[0058] The polarization SAR image classification method based on CNN and SVM is the same as that of Example 1-2, in which the repeated training of the AE network in step 5 to obtain the initial CNN convolutional layer parameters is performed according to the following steps:
[0059] 5a) Since the convolutional layer of CNN involves local perception areas, it is necessary to randomly select N×N image blocks in each training sample. The image block size selected in the present invention is 5×5;
[0060] 5b) Then use the image blocks selected in 5a) to train the AE network. The AE network has 3 layers: input layer, hidden layer, and output layer. Its working principle is to use the output layer to approximate the input layer to obtain another representation of the input layer characteristics , That is, the hidden layer feature; in the present invention, the number of convolution kernels of the CNN convolution layer is set to 20, which is the same as the number of neural units in the hidden layer of the AE network. The size of the convolution kernel is 5×5, which is consistent with the selected image block. The parameters between the input layer and the hidden layer of the trained AE network can be used as the parameters of the CNN convolutional layer.
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.
the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Similar technology patents

Monocular vision-based road detection method

Owner:XI'AN INST OF OPTICS & FINE MECHANICS - CHINESE ACAD OF SCI

Classification and recommendation of technical efficacy words

Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products