SAR image matching identification method
By estimating azimuth angles and fusion of feature data on SAR images, the matching recognition of SAR images is achieved by using SVM classifiers, which solves the problems of insufficient utilization of multi-dimensional features and slow operation speed in traditional methods, and improves the recognition accuracy and efficiency.
Patent Information
- Application Number
- CN202410613962.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-17
- Publication Date
- 2024-12-31
AI Technical Summary
Traditional SAR image matching recognition methods do not utilize multi-dimensional features, have low operating efficiency, low matching results accuracy, and are sensitive to lighting changes and occlusion.
By estimating the azimuth angle of the SAR image, the feature data of a variety of training samples within the angle threshold range were selected, and the feature data was fused by parallel strategies to obtain feature representation vectors, and image matching recognition was used by SVM classifiers.
The multi-dimensional feature utilization efficiency and accuracy of SAR image matching recognition can be improved, and the target recognition performance can be accurately identified while ensuring high running speed.
Smart Images
Figure CN119229159A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a SAR image matching and recognition method, which is suitable for classifying and recognizing SAR images by utilizing multi-dimensional features of SAR images. Background Art
[0002] SAR (Synthetic Aperture Radar) images are a type of remote sensing image acquired by a synthetic aperture radar system. Synthetic aperture radar uses a radar beam to scan the ground and record the echo signal, and then generates an image through signal processing technology. The SAR system acquires data by emitting pulsed electromagnetic waves and receiving the signals reflected back. Unlike optical remote sensing, SAR uses the characteristics of radar beams to acquire data in any weather conditions, even at night or under clouds. This makes SAR images very useful in surface observation and monitoring. SAR images have the following characteristics: high resolution, which can provide images with rich details; three-dimensional information, which can provide height information of the surface and generate digital elevation models of the surface; multi-frequency and polarization, which can use multiple frequencies and polarization modes for observation to obtain different types of information; reflectivity differences, which reflect the ability of surface objects to reflect radar waves; time series observation, which detects surface changes through continuous observation and comparison, etc.
[0003] Traditional feature matching recognition algorithms mainly include histogram matching algorithms and point feature extraction and matching algorithms.
[0004] The histogram matching algorithm is a commonly used SAR image object feature matching and recognition algorithm. The algorithm generates histogram features by statistically analyzing the grayscale distribution of pixels in the SAR image, and then achieves object matching and recognition by comparing the similarity between the histogram features of different objects. Common comparison methods include Euclidean distance, correlation coefficient, etc. Histograms with smaller distances or higher correlation coefficients are similar objects.
[0005] The advantages of the histogram matching algorithm are that it is simple, intuitive and easy to implement. It does not require complex mathematical operations and model fitting for SAR images, the extracted histogram features are impressive, and the calculation speed is fast. However, the histogram algorithm also has some limitations. First, this method can only perform statistics on the grayscale distribution of the entire image, and cannot capture spatial relationships and local changes. Second, the histogram can only provide one-dimensional feature information, and may not be able to accurately identify complex objects. In addition, the histogram features are sensitive to lighting changes and occlusion.
[0006] In summary, the histogram algorithm is a simple and effective SAR image object feature matching and recognition algorithm. Although it has limitations, it can be combined with other feature extraction methods to form a comprehensive feature descriptor to improve the accuracy and robustness of object recognition.
[0007] The point feature extraction and matching algorithm extracts key points and corresponding feature descriptors from the image, and then uses these descriptors to match and identify objects. The key points are found by detection algorithms such as Harris corner detection and FAST corner detection. Generally, local features such as texture, angle or edge in the image are selected as key points. After obtaining the key points, the feature descriptors are calculated based on the local areas around the key points. Commonly used descriptors include SIFT (scale invariant feature transform), SURF (speeded up robust features) and ORB (rotation invariant features). These descriptors are invariant to lighting, scale and rotation transformations, and can capture the unique characteristics of objects.
[0008] Feature matching is performed by comparing feature descriptors between different images and calculating the similarity or distance between features. Common matching algorithms include brute force matching, distance-based matching, and interest point-based matching. The matching threshold can be adjusted according to specific scenarios and applications.
[0009] The advantage of this algorithm is that it can extract local feature information in the image and has good robustness to scale, rotation and illumination changes. In addition, this algorithm does not rely on the grayscale distribution of the entire image, but only focuses on the local area near the key point, reducing the computational complexity and memory consumption. The disadvantage of the algorithm is that it may be affected by noise, occlusion and low texture, resulting in inaccuracy in key point extraction and descriptor calculation. The special properties of SAR images, such as strong scattering interference and complex background noise, also cause uneven distribution of key points, thus affecting the accuracy of matching. Summary of the invention
[0010] The technical problem solved by the present invention is: to overcome the deficiencies of the prior art and provide a SAR image matching and recognition method, which solves the problems that the traditional processing method does not make sufficient use of the multi-dimensional features in the SAR image, and the method has low operating efficiency, slow operating speed and low matching result accuracy.
[0011] The technical solution of the present invention is: a SAR image matching and recognition method, comprising:
[0012] Estimating the azimuth angle of the SAR image and screening out multiple training sample feature data within the angle threshold range;
[0013] A parallel strategy is used to fuse the selected feature data of multiple training samples to obtain a feature representation vector;
[0014] The obtained feature representation vector is used to train the SVM classifier, and the trained SVM classifier is used to realize the matching and recognition of SAR images.
[0015] The SAR image is estimated by azimuth angle, and multiple training sample feature data within the angle threshold range are screened out, including: based on the SAR image SVM classification and recognition method oriented to azimuth sensitivity, the target azimuth angle information obtained in the feature extraction stage is used to screen out the training feature data of the SVM classifier, specifically: the training data within the target azimuth angle range of ±10° and 180°±10° are selected respectively.
[0016] The training sample feature data includes one-dimensional distance direction, attribute scattering center and geometric features of the SAR image.
[0017] The parallel strategy is adopted to fuse the selected multiple training sample feature data to obtain the feature representation vector, including:
[0018] Different dimensional transformations are performed on the feature data of different training samples to obtain different feature data vectors. The mean-variance normalization method is used to convert the feature data vectors into a standard normal distribution with a mean of 0 and a standard deviation of 1. Then, the converted different feature data vectors are fused based on a parallel strategy to obtain a feature representation vector that can be sent to the SVM classifier for training.
[0019] The dimensional transformation processing is performed in different ways on different training sample feature data to obtain different feature data vectors, including:
[0020] Each attribute scattering center point of the SAR image is characterized by three elements, namely, the azimuth, the range coordinates, and the attribute scattering center amplitude. The number threshold of each element is set to N. If the number is insufficient, zero is added to unify the final vector length to 3N.
[0021] The one-dimensional range representation of the SAR image is to sum up the distances in the azimuth direction within the minimum circumscribed rectangle of the object and then average them to obtain a one-dimensional range feature vector. Then, the vector length is increased to 3N through cubic spline interpolation.
[0022] The length, width and aspect ratio of the SAR image are combined into a vector of length 3, and then repeated N times to generate a vector of length 3N.
[0023] The method of training an SVM classifier using the obtained feature representation vector and implementing matching and recognition of SAR images using the trained SVM classifier includes:
[0024] For binary classification, the SVM classifier maps the training data, i.e., the feature representation vector, to a high-dimensional linear feature space through a kernel function, constructs the optimal classification hyperplane in this linear space, and obtains the decision function of the classifier, and then classifies the feature representation vector;
[0025] For multi-classification, multiple SVM classifiers are used to define the decision boundaries between different categories, and multiple binary classifications are achieved to complete the multi-classification problem.
[0026] After the SVM classifier training is completed, the feature representation vector extracted from the SAR image is input into the SVM classifier to achieve matching and recognition of the SAR image.
[0027] For multi-classification, multiple SVM classifiers are used to define the decision boundaries between different categories, and multiple binary classifications are realized to complete the multi-classification problem, including: during training, samples of a certain category are classified into one category, and other samples are classified into another category. Then, k SVM classifiers are constructed for samples of k categories, and unknown samples are classified into the class with the largest classification function value during classification.
[0028] The kernel function is selected from one of a linear kernel, a polynomial kernel, a Gaussian kernel, and a Laplace kernel.
[0029] The advantages of the present invention compared with the prior art are:
[0030] The present invention is based on the basic principles and methods of feature fusion and support vector machine (SVM) classifier, and uses the azimuth estimation of the object to screen the training data, so as to introduce the feature matching recognition method of the object according to the multi-dimensional features of the object (mainly including geometric structure features and electromagnetic scattering features), and complete the classification and recognition of the object in the SAR image. Compared with the traditional processing method, the present method is suitable for multi-dimensional feature matching and recognition of objects in SAR images, while ensuring that the method has high operating efficiency, speed and high-precision matching object capabilities, and can excellently complete the object classification and identity recognition tasks. The present invention is a target recognition method based on a combination of multiple classifiers, while traditional SAR image target classification and recognition usually uses a single classifier for recognition, which has good recognition performance on some sample features, but poor recognition performance on other sample features. Since for the same recognition task, the samples misclassified by different classifiers are not exactly the same, that is, there is a certain complementarity between different classifiers. Therefore, the present method uses a combination of multiple classifiers in SAR image matching and recognition to significantly improve the target recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 Schematic diagram of the object length characteristics changing with azimuth angle.
[0032] Figure 2 Schematic diagram of the object mean feature changing with azimuth angle.
[0033] Figure 3 Schematic diagram of the object variance coefficient characteristics changing with azimuth angle.
[0034] Figure 4Schematic diagram comparing one-dimensional distance before interpolation (left picture) and after interpolation (right picture).
[0035] Figure 5 Schematic diagram of the mapping relationship between the input space and the high-dimensional feature space.
[0036] Figure 6 Schematic diagram of the IOU calculation process. DETAILED DESCRIPTION
[0038] The present invention includes training data screening, feature fusion and matching recognition. First, the SAR image is estimated in azimuth, and multiple training sample feature data within the angle threshold range is screened; the multiple screened training sample feature data are fused using a parallel strategy to obtain a feature representation vector; the obtained feature representation vector is used to train an SVM classifier, and the trained SVM classifier is used to achieve matching recognition of SAR images. The specific process is as follows:
[0039] 1. Training data screening
[0040] Since SAR images reflect the structural characteristics of the object, especially at locations with basic scattering mechanisms such as dihedral angles and trihedral angles, the SAR echo is strong and appears as a strong scattering point in the image. These basic scattering mechanisms are highly sensitive to the incident direction of the incident wave, and the echo intensity has obvious fluctuation characteristics under different incident directions. Therefore, the SAR image of the object has a strong azimuth sensitivity. The characteristic value of the object will change with the different azimuth angles of the object.
[0041] The geometric scale features of 698 object images of three types of ground objects, BMP-2, BTR-70 and T72, in the MSTAR data were extracted. The extracted geometric scale features include length, aspect ratio, area, shape complexity, mean and variance coefficient. The azimuth angles corresponding to these 698 object images were estimated to obtain the azimuth angle information. The azimuth angles were arranged in sequence from 0° to 360° with the azimuth angle as the x-axis, and the features corresponding to the azimuth angles were marked on the image in the form of scatter plots. The relationship between the features of the training samples and the azimuth angles is shown in the figure below: Figure 1 , Figure 2 , Figure 3 shown.
[0042] It can be clearly seen from the figure of the relationship between features and azimuth that, on the one hand, the feature value of the object will change with the different azimuth angles of the object, indicating that the object SAR image has obvious azimuth sensitivity; on the other hand, the distribution of the object feature value presents a 180° fuzzy phenomenon, that is, objects with an azimuth difference of 180° have similar feature values.
[0043] Therefore, the SAR image object SVM classification and recognition algorithm for azimuth sensitivity is adopted. First, the object azimuth information obtained in the feature extraction stage is used to screen the training samples of the SVM classifier. When screening the training samples, considering the 180° fuzzy phenomenon of the object azimuth extraction, the training data within the range of object azimuth ±10° and 180°±10° are selected respectively. For example, if the azimuth of the current test sample is 45°, the training sample feature data within the range of [35°, 55°] and [215°, 235°] are selected to train the SVM classifier.
[0044] 2. Feature Fusion
[0045] Feature fusion is to use a certain strategy to generate a new feature representation vector from the original multiple feature vectors (one-dimensional distance, attribute scattering center, geometric features). At present, the commonly used fusion strategies are serial, parallel and genetic algorithms. Since the dimensions of multiple feature vectors are quite different, a parallel strategy is adopted. The specific principles are as follows:
[0046] Assume that A and B are two feature spaces defined in the pattern sample space Ω. For any sample Γ∈Ω, the corresponding feature representation vectors are α∈A and β∈B. The parallel fusion strategy combines these two feature representation vectors into a vector γ, that is,
[0047] γ=α+i·β
[0048] Where i is the fusion coefficient. Since the parallel fusion strategy requires the dimension of the feature vector to remain consistent, different features are processed in different ways:
[0049] Each attribute scattering center point is characterized by three elements, namely its azimuth, distance coordinates, and attribute scattering center amplitude. The number threshold of attribute scattering centers is set to 60. If the number of extracted attribute scattering centers exceeds the threshold, only the 60 with the highest attribute scattering center amplitudes are taken. If there are less than 60, the length of the vector is unified to 60×3=180 by zero padding, which is recorded as V1.
[0050] The one-dimensional distance represents the sum of the distance and azimuth in the smallest circumscribed rectangle of the object, and then the average is obtained to obtain the one-dimensional distance feature. The length of this vector is usually between 20 and 40. Through cubic spline interpolation, the vector length is increased to 180, which is recorded as V2. Figure 4 It can be seen that the one-dimensional distance curve is smoother after interpolation, and the vector length is consistent with other features.
[0051] Combine the length, width, and aspect ratio of the object into a vector of length 3, and repeat 60 times to generate a vector of length 180, recorded as V3.
[0052] Before feeding the data into SVM, it is usually necessary to perform normalization (also called feature scaling) to ensure that the numerical ranges of different features are consistent. Mean-variance normalization is used to convert feature data V1, V2, and V3 into a standard normal distribution with a mean of 0 and a standard deviation of 1. Then, different features are fused based on a parallel strategy to obtain a vector V:
[0053] V=λ1V1+λ2V2+λ3V3
[0054] Finally, the fused feature vector V is sent to the SVM classifier for matching and recognition.
[0055] 3. Matching and identification
[0056] SVM classifier, compared with traditional statistics, the SVM algorithm is not based on the traditional empirical risk minimization principle, but on the principle of structural risk minimization (SRM). It can well solve the problem of constructing high-dimensional models with a limited number of samples, and the constructed model has good predictive performance.
[0057] For the nonlinear binary classification problem, assume that the sample set {(x i ,y i ),i=1,2,...,l}, consisting of two categories, if x i ∈R (N) Belongs to category 1, then marked as positive (y i =1), if it belongs to the second category, it is marked as negative (y i =-1). The object of learning is to construct a decision function that can classify the test data as correctly as possible.
[0058] like Figure 5 As shown in Figure 1, when the training set is nonlinear, the training set data x is mapped to a high-dimensional linear feature space through a nonlinear function φ(·), and the optimal classification hyperplane is constructed in this linear space whose dimension may be infinite, and the decision function of the classifier is obtained.
[0059] Therefore, in the nonlinear case, the classification hyperplane is
[0060] ω T φ(x)+b=0
[0061] The decision function is
[0062]
[0063] The φ(·) function is generally called the kernel function. Common kernel functions include linear kernel, polynomial kernel, Gaussian kernel, Laplace kernel, etc.
[0064] The above specific assumptions may not be completely useful in practice, because the training data may not be completely separated by the hyperplane. In this case, a slack variable is introduced, denoted by ξ i , leaving a soft margin to contain the possibility of violation. Each sample has a corresponding slack variable that characterizes the degree to which the sample does not satisfy the constraint. This variable is specified as a number greater than 0, as shown below:
[0065]
[0066] By introducing the Lagrange multiplier α i and β i Solving the optimization problem with inequality constraints, we can obtain:
[0067]
[0068] Then SVM needs to find a linear separation hyperplane with the maximum margin in this high-dimensional space. C>0 is the penalty parameter of the error term. In addition, K(x i ,x j )=φ(x i ) T φ(x j ) is called a kernel function. One of the commonly used kernel functions, the Gaussian kernel function, is as follows:
[0069] K(x i ,x j ) = exp(-γ||x i -x j || 2 ),γ>0
[0070] After solving, we get the classification decision function:
[0071]
[0072] The object function is:
[0073]
[0074] When the SVM algorithm is used for multi-classification problems, it essentially uses multiple SVM classifiers to define the decision boundaries between different categories. Implementing multiple binary classifications means completing the multi-classification problem. There are two specific implementations. The first is to classify samples of a certain category into one category and other samples into another category during training. In this way, k SVMs are constructed for samples of k categories. During classification, unknown samples are classified into the category with the largest classification function value. The second is to design an SVM between any two categories of samples. This method is needed in the classification of k categories. After testing each SVM classifier separately, the scores of each category are accumulated, and the category corresponding to the highest score is selected as the category of the test data.
[0075] In the classification of SAR images based on SVM, the support vector (i.e., fused feature vector) used for training comes from the features extracted from the object, such as geometric features such as object length, width, aspect ratio, and electromagnetic scattering features such as one-dimensional range image and object attribute scattering center information. The extracted features are then input into the SVM machine learning classifier to complete the determination of the object category.
[0076] In addition, whether the object is matched and recognized successfully needs to consider two aspects: whether the ROI area of the object is correctly extracted and whether the category of the object is correctly identified. Only when the object area is correctly extracted, the identification of the object type is effective.
[0077] Set the IOU threshold. When the IOU of the real label and the judgment result is greater than the threshold, it can be considered that the ROI area of the object is extracted correctly. Figure 6 The IOU shown is calculated by dividing the intersection of the true label and the judgment result by the union.
[0078] The matching and recognition of object types are evaluated using accuracy, precision, and recall.
[0079] Accuracy is used to measure the proportion of all 35 object samples that are correctly classified by the classification model. It is the ratio of the number of correctly classified samples divided by the total number of samples, usually expressed as a percentage. The higher the accuracy, the better the performance of the model. The calculation method is
[0080]
[0081] Based on different IOU thresholds, different accuracy evaluation indicators can be obtained. 50 The IOU threshold is 0.5, and the average accuracy of all samples in a certain category is calculated. 50:95 The IOU threshold is set from 0.5 to 0.95 with a step size of 0.05 for a total of 10 values, and the average accuracy is calculated.
[0082] Precision is used to evaluate the performance of classification models. It measures the proportion of samples that are actually positive among those predicted by the model as positive. The precision value ranges from 0 to 1, where 1 means that all predicted positive classes are correct. The calculation method is:
[0083]
[0084] Recall, also known as sensitivity or true positive rate, is an indicator for evaluating the performance of a classification model. It measures the proportion of samples that the model successfully predicts as positive among all samples that are actually positive. The recall value ranges from 0 to 1, where 1 means that the model successfully captures all positive samples. The calculation method is:
[0085]
[0086] Among them, TP (True Positive) is the positive sample predicted by the model as the positive class, that is, the sample predicted correctly. FP (False Positive) is the negative sample predicted by the model as the positive class, that is, the sample predicted incorrectly. FN (False Negative) is the positive sample predicted by the model as the negative class, that is, the sample that was not predicted.
[0087] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solution of the present invention by using the technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.
Claims
1. A SAR image matching and recognition method, characterized in that: include: Estimating the azimuth angle of the SAR image and screening out multiple training sample feature data within the angle threshold range; A parallel strategy is used to fuse the selected feature data of multiple training samples to obtain a feature representation vector; The obtained feature representation vector is used to train the SVM classifier, and the trained SVM classifier is used to realize the matching and recognition of SAR images.
2. The SAR image matching and recognition method according to claim 1, characterized in that: The SAR image is estimated by azimuth angle, and multiple training sample feature data within the angle threshold range are screened out, including: based on the SAR image SVM classification and recognition method oriented to azimuth sensitivity, the target azimuth angle information obtained in the feature extraction stage is used to screen out the training feature data of the SVM classifier, specifically: the training data within the target azimuth angle range of ±10° and 180°±10° are selected respectively.
3. The SAR image matching and recognition method according to claim 1, characterized in that: The training sample feature data includes one-dimensional distance direction, attribute scattering center and geometric features of the SAR image.
4. The SAR image matching and recognition method according to claim 3, characterized in that: The parallel strategy is adopted to fuse the selected multiple training sample feature data to obtain the feature representation vector, including: Different dimensional transformations are performed on the feature data of different training samples to obtain different feature data vectors. The mean-variance normalization method is used to convert the feature data vectors into a standard normal distribution with a mean of 0 and a standard deviation of 1. Then, the converted different feature data vectors are fused based on a parallel strategy to obtain a feature representation vector that can be sent to the SVM classifier for training.
5. The SAR image matching and recognition method according to claim 4, characterized in that: The dimensional transformation processing is performed in different ways on different training sample feature data to obtain different feature data vectors, including: Each attribute scattering center point of the SAR image is characterized by three elements, namely, the azimuth, the range coordinates, and the attribute scattering center amplitude. The number threshold of each element is set to N. If the number is insufficient, zero is added to unify the final vector length to 3N. The one-dimensional range representation of the SAR image is to sum up the distances in the azimuth direction within the minimum circumscribed rectangle of the object and then average them to obtain a one-dimensional range feature vector. Then, the vector length is increased to 3N through cubic spline interpolation. The length, width and aspect ratio of the SAR image are combined into a vector of length 3, and then repeated N times to generate a vector of length 3N.
6. The SAR image matching and recognition method according to claim 1, characterized in that: The method of training an SVM classifier using the obtained feature representation vector and implementing matching and recognition of SAR images using the trained SVM classifier includes: For binary classification, the SVM classifier maps the training data, i.e., the feature representation vector, to a high-dimensional linear feature space through a kernel function, constructs the optimal classification hyperplane in this linear space, and obtains the decision function of the classifier, and then classifies the feature representation vector; For multi-classification, multiple SVM classifiers are used to define the decision boundaries between different categories, and multiple binary classifications are achieved to complete the multi-classification problem; After the SVM classifier training is completed, the feature representation vector extracted from the SAR image is input into the SVM classifier to achieve matching and recognition of the SAR image.
7. A SAR image matching and recognition method according to claim 6, characterized in that: For multi-classification, multiple SVM classifiers are used to define the decision boundaries between different categories, and multiple binary classifications are realized to complete the multi-classification problem, including: during training, samples of a certain category are classified into one category, and other samples are classified into another category. Then, k SVM classifiers are constructed for samples of k categories, and unknown samples are classified into the class with the largest classification function value during classification.
8. The SAR image matching and recognition method according to claim 4, characterized in that: The kernel function adopts one of a linear kernel, a polynomial kernel, a Gaussian kernel, and a Laplace kernel.