CN117876804A - SAR target subspace feature optimization method - Google Patents

Abstract

The invention belongs to the technical field of radar remote sensing application, and particularly relates to a SAR target subspace feature optimization method. The method of the invention optimizes the mode features extracted from the SAR image, and utilizes the unique feature attributes of the SAR image to restrict the optimizing process, thereby realizing the optimization of the feature subspace level and obtaining the feature subspace with the highest target recognition accuracy. Therefore, the SAR target recognition accuracy is improved while the dimension of the feature is reduced, and the optimal feature subset is obtained to reconstruct subspace, so that the SAR target recognition accuracy has excellent interpretability.

Description

SAR target subspace feature optimization method

Technical Field

The invention belongs to the technical field of radar remote sensing application, and particularly relates to a SAR target subspace feature optimization method.

Background

The synthetic aperture radar (Synthetic Aperture Radar, SAR for short) has the characteristics of all-weather, all-weather and the like, and is an important earth observation means. SAR image target recognition utilizes the information characteristics of SAR images to realize the discrimination and classification of targets, and is widely applied to the fields of ship monitoring, exploration resources, target positioning and the like.

The accuracy of SAR image target recognition has close relation with feature extraction and feature selection. However, with the development of the SAR imaging technology, the characteristic information of the SAR image is exploded, which brings great challenges to tasks such as classification and detection. There are a large number of redundant or uncorrelated features in the data that increase computational costs, or may cause the model to suffer from over-fitting phenomena that reduce the performance and generalization ability of the learning model, such phenomena often being referred to as "dimension disasters". Therefore, it is necessary to search for better feature extraction, feature dimension reduction, feature selection and other feature processing modes to obtain better performance in SAR target recognition.

In the field of target recognition, feature extraction generally includes mode feature extraction such as Principal Component Analysis (PCA), texture feature matrix, information entropy and the like, and feature extraction by using a deep network mode. The feature dimension reduction method generally uses a mapping or conversion method to reduce the dimension of original data, so that data with the same category are similar and data with different categories are separated. The feature selection method is also actually intersected with the dimension reduction process. The original data is measured by using the measurement standard, and a feature subset with lower dimension is selected. For high-dimensional data, evaluating the superiority of a given feature set under a certain evaluation criterion, retaining the superior features therein, and deleting irrelevant and redundant features. However, these methods generally have good effects on optical images, but the SAR image imaging mechanism is different from that of the optical images to some extent, so that these methods cannot be directly used in the SAR image target recognition field, which makes the feature dimension reduction and selection scheme of the SAR image difficult.

The problems that the extracted characteristic effect is poor, the characteristic of the SAR image is not considered in the selection process, the combination relation and weight distribution among the characteristics are not considered, the interpretability of the characteristic selection on the SAR target is not strong and the like exist in the use of the traditional characteristic selection scheme in the SAR image target recognition field, and the characteristics are mainly expressed in the following four aspects:

(1) The feature extraction method selected for the optical image does not necessarily have a good effect on the SAR image, and characteristics of the SAR image are not considered.

(2) Common evaluation criteria adopted in the process of selecting general features include distance measurement criteria, information measurement criteria, relevance measurement criteria, consistency measurement criteria and the like, and the common evaluation criteria tend to focus on the features themselves, but the features cannot be selected by taking task demands as guidance, so that optimal features extracted from a model cannot necessarily achieve optimal results on classification tasks.

(3) For the feature selection scheme in the traditional target recognition process, the final output result is a group of feature combinations, the group of features are sent to the subsequent decision process in an average weighting mode, and the individual feature weight problem of the optimal feature combinations is not considered.

(4) For the traditional feature extraction process, especially the depth feature extraction process, although the recognition accuracy can be improved after feature selection, the method has poor interpretability. For most of the fields to which SAR images are applied, having interpretability is essential for subsequent research.

Disclosure of Invention

The invention aims to solve the problems and the defects that the extracted characteristic effect is poor, the characteristic of SAR images is not considered in the selection process, the combination relation and weight distribution among the characteristics are not considered, the interpretation of the characteristic selection on SAR targets is not strong, and the like in the use of the SAR image target recognition field, and particularly provides an SAR target subspace characteristic optimization selection method based on an improved genetic algorithm and an integrated classifier.

The technical scheme of the invention is as follows:

a SAR target subspace feature optimization method comprising the steps of:

s1, acquiring data from a known SAR image dataset, and preprocessing to obtain a training dataset;

s2, PCA principal component analysis feature extraction is carried out on the training data set, and the training data set is defined as X epsilon R ^m×n Each sample ism represents the number of samples, n represents the dimension of the training data set, and the custom parameter k represents the extraction dimension (for example, the value 100), specifically including:

a1, calculating the average value of each dimension:

a2, subtracting the obtained average value from each dimension to obtain a matrix, and carrying out centering treatment on input data:

a3, constructing a covariance matrix:

a4, decomposing the eigenvalue of the covariance matrix C, and taking the eigenvectors corresponding to the maximum k eigenvalues to form a dimension-reducing matrix V epsilon R ^k×n ；

a5, multiplying the dimension-reducing matrix V by the original matrix to reduce the dimension, so as to obtain a dimension-reduced matrix, namely Y' =XV ^T I.e., the PCA feature set after dimension reduction (e.g., selecting the first 100-dimensional features);

performing GLCM gray scale formula matrix calculation on SAR images in a training data set to obtain contrast, dispersion, homogeneity, energy, correlation and ASM attributes, forming a 6-dimensional feature serving as a constraint condition, and presetting a threshold according to the dimension of the 6-dimensional feature;

s3, performing feature selection by using an improved genetic algorithm, wherein the improved method is to restrict the intersecting process of the genetic algorithm by using the constraint condition obtained in the S2, and the constraint method is to restrict the genetic algorithm to optimize the extracted PCA features by using SAR target characteristics if 6-dimensional constraint features between the original SAR image and the image with the optimized post-feature reconstruction are smaller than a preset threshold, and if not, intersecting father individuals and mother individuals in the genetic algorithm. Taking the 100-dimensional PCA feature as an example, where different feature cross-combinations correspond to different subsets of the feature set, and also correspond to solving for each subspace in the space.

S4, training the classifier by using the obtained feature subsets to obtain a trained classifier;

s5, inputting the obtained SAR image into a trained classifier to obtain the identification accuracy and the final optimal feature subspace.

The SAR target subspace feature optimization method based on the improved genetic algorithm and the integrated classifier has the advantages that on the basis of traditional feature dimension reduction and feature selection, the SAR target subspace feature optimization method based on the improved genetic algorithm and the integrated classifier is provided. The method has the important innovation point that the optimization process is constrained by utilizing the unique characteristic attribute of the SAR image, the optimization of the characteristic subspace level is realized, in the process, the characteristic dimension required by classification is reduced, the SAR target recognition accuracy is improved, meanwhile, the optimal characteristic subspace obtained by the algorithm can be subjected to sample reconstruction, and the key characteristic part in the SAR target recognition process is explained.

Drawings

FIG. 1 is a block flow diagram of a preferred method of SAR target subspace characterization based on an improved genetic algorithm and integrated classifier of the present invention.

Fig. 2 is a SAR image of a target including a ship, an airplane, an armored car, and the like.

Fig. 3 is a flow chart of a feature reconstruction sample after method selection.

Detailed Description

The invention aims to improve the accuracy of SAR image target recognition and the interpretability of key features, so that an improved genetic algorithm based on SAR characteristic constraint is applied to an integrated machine learning classifier, and other mainstream classifiers can be adopted.

The specific embodiments are as follows:

1. first, four SAR image datasets Mstar, openSARShip, T, SAR-AIRcraft-1.0 are resized, and the sizes of the four datasets are respectively 158×158, 100×100, 70×70, 199×175. As shown in fig. 2, the experimental data set contains rich types of vessels, armored vehicles, aircraft, and the like.

2. The four data sets were partitioned into training samples and test samples according to the distribution of the number of data sets and the number of test sets in table 1:

table 1SAR target data set used for quantitative evaluation

Data set	Training set number	Number of test sets	Number of categories
				Mstar	2747	2425	10
OpenSARShip	480	120	3
				T72	2388	2189	8
SAR-AIRcraft-1.0	13168	3298	7

3. And respectively carrying out PCA principal component analysis on the data set to extract features, storing the front 100-dimensional features, then carrying out GLCM gray formula matrix calculation on the SAR image to obtain 6 features such as contrast (const), dispersion (dissimilarity), homogeneity (homogeny), energy (energy), correlation (ASM) Attribute (ASM) and the like, and taking the feature dimensions as constraint conditions of an optimization algorithm.

4. And (3) setting a threshold according to the dimension of the constraint features extracted from different data sets, constraining the crossing process of the basic genetic algorithm, if the constraint is smaller than the set threshold, crossing a father individual and a mother individual in the genetic algorithm, otherwise, not crossing, and constraining the selection process of the feature subspace of the genetic algorithm by utilizing SAR target characteristics. And meanwhile, carrying out parameter adjustment on the genetic algorithm, and setting the cross probability of 0.7, the variation probability of 0.1, the population number of 50, the iteration number of 1000 and the fitness function as classification accuracy.

5. Parameters are set for the integrated machine learning classifier based on the decision tree, and the subsampled proportion of the weight training is set to be 0.8, the learning rate is 0.05, each leaf is at a node 96, and the depth of the decision tree is 5.

6. And sending the training set into a training process, screening out the optimal feature subset and verifying on the test set. Meanwhile, the optimal feature subset can reconstruct a sample and compare the original SAR sample, and the mode is shown in figure 3. The feature dimension and classification accuracy of the data set before and after the method of the invention are compared as shown in table 2:

table 2 comparison table of front-to-back accuracy for SAR target data set using the method

The method mainly optimizes the mode features extracted from the SAR image, utilizes the unique feature attributes of the SAR image to restrict the optimization process based on the genetic algorithm, realizes the optimization of the feature subspace level, obtains the feature subset which can lead the target recognition accuracy to reach the highest, and has excellent interpretability on subspace reconstruction by the optimal feature subset. Meanwhile, the optimization method provided by the invention reduces the dimension of the feature dimension, improves the SAR target recognition accuracy, and provides a feasible scheme for subsequent practical application deployment.

Claims (1)

1. A SAR target subspace feature optimization method, comprising the steps of:

s1, acquiring data from a known SAR image dataset, and preprocessing to obtain a training dataset;

s2, respectively performing PCA principal component analysis feature extraction on the training data set, and defining the training data set as X epsilon R ^m×n Each sample ism representsThe number of samples, n, represents the dimension of the training data set, and the custom parameter k represents the feature extraction dimension; the method specifically comprises the following steps:

a1, calculating the average value of each dimension:

a2, subtracting the obtained average value from each dimension to obtain a matrix, and carrying out centering treatment on input data:

a3, constructing a covariance matrix:

a4, decomposing the eigenvalue of the covariance matrix C, and taking the eigenvectors corresponding to the maximum k eigenvalues to form a dimension-reducing matrix V epsilon R ^k×n ；

a5, multiplying the dimension-reducing matrix V by the original matrix to reduce the dimension, namely Y' =XV ^T Obtaining a PCA feature set after dimension reduction;

performing GLCM gray scale formula matrix calculation on SAR images in a training data set to obtain contrast, dispersion, homogeneity, energy, correlation and ASM attributes, and forming a 6-dimensional characteristic as a constraint condition;

s3, performing feature selection by using an improved genetic algorithm, wherein the improved method is to restrict the intersecting process of the genetic algorithm by using the constraint condition obtained in the S2, and the constraint method is to restrict the genetic algorithm to optimize the extracted PCA feature set by using SAR target characteristics if 6-dimensional constraint features between an original SAR image and an image with optimized post-feature reconstruction are smaller than a preset threshold, and a father individual and a mother individual in the genetic algorithm can intersect, otherwise, the father individual and the mother individual can not intersect;

s4, training the classifier by using the obtained feature subsets to obtain a trained classifier;

s5, inputting the obtained SAR image into a trained classifier to obtain the identification accuracy and the final optimal feature subspace.