CN106485239A - One kind is using one-class support vector machines detection river mesh calibration method - Google Patents

Abstract

In order to solve the problems in the prior art that it is impossible to learn through training samples and the detection results are not accurate enough, the present invention provides a method for detecting river targets using a single-class support vector machine, extracting spectral features in remote sensing images and then going through a rough screening process. Based on the spectral features, the river candidate regions are extracted, and then the feature vectors generated by the spectral values are used to train the one-class support vector machine. Based on the results of the coarse screening process, fine detection is performed to segment the image of the river candidate area, generate shape features, and finally realize target detection according to the shape index threshold. The invention uses the single-class support vector machine method to only need one class of training samples, thereby improving the detection efficiency; meanwhile, the detection accuracy is improved through two steps of coarse screening and fine detection.

Description

A Method of Detecting River Objects Using One-Class Support Vector Machine

technical field

The invention relates to the fields of image processing and computer vision, in particular to a method for detecting river targets in remote sensing images using a single-class support vector machine method.

Background technique

Target detection and tracking is one of the popular research directions in the field of image processing and computer vision. It has important research value in military imaging guidance and tracking military targets, as well as civil security monitoring and intelligent human-computer interaction.

Among the existing detection methods, the difficulty of target detection research lies in the effective representation of the target, and the target matching problem caused by changes in the target scale, rotation angle, illumination, etc. caused by various reasons. The characterization of the target determines the matching algorithm to a certain extent, including: using local contour features to represent the target, first extracting the contour of the image target, and filtering out the noise edge by automatically selecting a threshold to obtain a salient contour, which can effectively reduce the number of contour segments. Reduce the time and space complexity of subsequent feature extraction and the entire detection process. In order to overcome changes in the scale and rotation of the target, Lowe proposed the Scale Invariant Feature Transform (SIFT). By calculating the multi-scale Gaussian difference image and finding the local maximum point, the scale and rotation invariant features within a certain range are obtained.

In object detection, in addition to the characteristics of the object itself, contextual constraints can also be used to achieve effective detection of image objects. In remote sensing image target detection, due to the large size and complexity of the image, in order to avoid more false alarms, it is necessary to take certain measures to remove false alarm targets as much as possible. The general method is to screen candidate target areas by simple features first, and then fine-tune the candidate areas. Furthermore, for the detection of aircraft targets in remote sensing images, a cascade classifier is used to detect candidate target windows, and then the Hough Forest algorithm is used to make a second judgment on the candidate windows to filter out false alarms, which improves the detection efficiency and saves operation time.

Commonly used classification methods such as K-nearest neighbor classifier and Bayesian classification method all have the disadvantage of complicated process.

For example, K-Nearest Neighbor is the most popular and easy-to-understand classifier algorithm. Calculate the distance between the test sample and each training sample, take the smallest K of them, and vote according to the marks of the K training samples to get the test sample. mark. The idea of the algorithm is clear and simple. However, the amount of calculation is too large for massive data, and each training sample has a distance that must be measured, which consumes a lot of time.

The Bayesian classification method needs to use multiple stages to complete the calculation: the first stage - the preparation stage, the task of this stage is to make the necessary preparations for the Naive Bayesian classification, the main work is to determine the feature attributes according to the specific situation, and to Each feature attribute is properly divided, and then a part of the items to be classified is manually classified to form a training sample set. The input of this stage is all the data to be classified, and the output is feature attributes and training samples. This stage is the only stage that needs to be completed manually in the entire Naive Bayesian classification, and its quality will have an important impact on the entire process. The quality of the classifier is largely determined by the feature attributes, feature attribute division, and training sample quality.

The second stage - classifier training stage, the task of this stage is to generate a classifier, the main work is to calculate the frequency of occurrence of each category in the training samples and the conditional probability estimation of each category for each feature attribute division, and The results are recorded. Its input is feature attributes and training samples, and the output is a classifier. This stage is a mechanical stage, which can be automatically calculated by the program according to the formula discussed above.

The third stage - the application stage. The task of this stage is to use the classifier to classify the items to be classified, the input is the classifier and the items to be classified, and the output is the mapping relationship between the items to be classified and the categories. This stage is also a mechanical stage, completed by the program.

Rivers are important targets in remote sensing images, and river target detection is widely used in military and civil affairs. In the prior art, in the target detection process, various features such as color (spectral) features, shape structure features, texture features, context features, SIFT features, etc. are often used in image target detection. In the actual target detection application, the principle of feature selection is to effectively characterize the target. For different targets, the feature selection results are also different.

The spectral feature is the gray value of the image pixel, which represents the spectral reflection characteristics of the ground object. It is often used as the basis for the classification of the ground object, but it cannot reflect the geometric structure characteristics of the target to be detected. The image patches generated after the original image is segmented have shape and structure characteristics, and a covariance matrix is constructed based on the vector composed of pixel coordinates in the image patches, and the aspect ratio, shape index, density, and main parameters of the image patches can be extracted further. characteristics such as direction. Feature extraction based on image segmentation is greatly affected by the segmentation method and results, so it has high uncertainty. There are many texture representation methods, the most commonly used is the texture feature description based on the gray level co-occurrence matrix. The gray level co-occurrence matrix is defined by the joint probability density of pixels at two positions. It not only reflects the distribution characteristics of brightness, but also reflects the distribution characteristics of the positions between pixels with the same brightness or close to brightness. It is related to the brightness change of the image. The second-order statistical features of are the basis for defining a set of texture features. Texture features are often not very distinguishable for general target ground targets. Commonly used texture features include homogeneity, contrast, energy, and entropy.

In a word, in the existing technology: 1. The features used in the remote sensing image river target detection method include color (spectral) feature, shape structure feature, texture feature, context feature, SIFT feature, etc., the algorithm is complex, and the detection result is not accurate enough.

2. In the target detection task, it is required to detect a specific target category from the background, but the number of categories contained in the background is uncertain, and it is generally impossible to automatically learn the categories in the background.

Contents of the invention

In order to solve the problems in the prior art that the training samples cannot be used for learning and the detection results are not accurate enough, the present invention provides a method for detecting river targets using a single-class support vector machine. The classification vector machine method overcomes the difficulty of selecting non-target category samples in target detection, simplifies the learning stage in the process of target detection, and improves detection efficiency; at the same time, the accuracy of detection is improved through two links of coarse screening and fine detection.

The technical solution adopted by the present invention to solve the problem is: adopt the following steps:

A. Select the band: Select the band combination in the remote sensing image to distinguish the feature and extract the river, so as to identify the water body, vegetation and other ground objects;

B. Feature analysis and selection: Select the image processed in step A, and perform spectral feature analysis: the spectral feature used in spectral analysis is: rivers and other water targets in remote sensing images, distinguish river targets by gray value;

C. Rough screening process: based on spectral features, extract river candidate areas: select several target samples, extract the spectral value of each pixel as a classification feature, and use the formula x _i = (r _i , g _i , b _i ) to generate a feature vector, Use the generated feature vector to train the single-class support vector machine, and use the RBF kernel function during training Carry out training, and determine the classification model through 10-fold cross-validation, perform single-class classification on the entire remote sensing image, and obtain the water body category extraction result; among them, r _i corresponds to the red component; g _i corresponds to the green component; b _i corresponds to the blue component; | | _xi -x _j || represents the Euclidean distance between any two points x _i and x _j in the space; γ is the kernel parameter.

D. Fine detection process: Based on the results in the C link, image segmentation is performed on the river candidate area to generate shape features: On the basis of the results obtained in the C link rough screening, the image segmentation technology is used to set the threshold parameter, which will be greater than the threshold and Adjacent pixels are merged to generate target category connected regions of different sizes, and then the shape feature index is calculated from the pixels in each connected region And set the area threshold to remove the small area, and finally merge the image of the removed small area into the background area; where, the border length e: the number of border pixels, the border length of a pixel is 1; area A: the composition of the The total number of pixels of the object, where the length of a pixel edge is set to 1.

E. Target detection: set the shape index threshold, and determine the river target according to the shape feature index obtained in the D link, and realize the target detection.

The fine detection process of the D link: using the threshold segmentation method, using the area A as the threshold, extracting the target candidate area, and generating the shape feature S at the same time.

Further, in the band selection link, the 4, 3, and 2 bands of the remote sensing image are selected, and red, green, and blue are respectively assigned to distinguish ground features.

Further, the gray value of the feature analysis and selection link is 10-20.

Further, the threshold parameter for image segmentation in the fine detection process is: Th=10; the area threshold is: AT=50.

Further, the shape index threshold in the target detection link is ST=2.5.

The beneficial effects of the present invention are: the present invention only needs one class of training samples using the single-class support vector machine method, overcomes the difficulty of selecting non-target class samples in target detection, simplifies the learning stage in the target detection process, and improves detection efficiency ; At the same time, the accuracy of the detection is improved through the two links of coarse screening and fine detection.

Description of drawings

Figure 1 is a schematic diagram of a remote sensing image.

Figure 2 is a schematic diagram of the river candidate regions determined by the single-class support vector machine method.

Figure 3 is a schematic diagram of river candidate regions determined by the nearest neighbor classification method.

Fig. 4 is a schematic diagram of river candidate areas determined by Bayesian classification method.

Figure 5 is a schematic diagram of removing small areas smaller than 50 pixels.

Figure 6. Schematic diagram of matching results when the shape feature index is greater than 2.5.

detailed description

As shown in Figures 1 and 2, a method for detecting river targets using a single-class support vector machine adopts the following steps:

A. The band selection of the remote sensing image, select the band combination suitable for the distinction of ground features and river extraction: select the 4, 3, and 2 bands of the remote sensing image, and assign red, green, and blue colors respectively, and the combination of these bands forms a false color image. The images of ground features are rich, clear and well-layered, and can better identify water bodies and vegetation.

B. Feature analysis and selection: Select the image processed in step A to perform spectral feature analysis: the spectral feature used in spectral analysis is a prominent feature of water objects such as rivers: in remote sensing images, river targets have lower gray values 10-20.

C. Rough screening process: Based on the spectral analysis of the gray features in the B link, extract the candidate area of the river: first, select 1000 pixels as the target category samples, and extract the spectral value of each pixel as the classification feature x _i =(r _i , g _i , b _i ), generate 1000 feature vectors x _i , where _{ri is the red component, g i is} the green component, and b _i is the blue component. A one-class support vector machine is trained using these feature vectors. Given l training samples, the single-class support vector machine optimization problem is solved, namely

And satisfy (w ^T φ( _xi )+b)≥1-ξ _i , ξ _i ≥0 (2).

Among them, C>0 is the penalty coefficient, φ( _xi ) is the function that maps the vector _xi to a high-dimensional space; w is the normal vector of the classification surface in the feature space, w ^T is the transpose of w, and b is the The intercept, _ξi is the relaxation factor.

Generally, K(x _i , x _j )≡φ(x _i ) ^T φ(x _i ) is the kernel function in the support vector machine.

In the training process, the feature vector and the category label are combined together as the training input parameters, that is, ( _xi , y _i ), where y _i is the category label, and in this classification method, y _i =1, that is, the category All are positive. Choose RBF kernel function in training Among them, || _xi -x _j || represents the Euclidean distance between any two points x _i and x _j in the space, and the classification parameters are obtained through 10-fold cross-validation. The 10-fold cross-validation method is to randomly divide the 1000 training samples into 10 equal subsets, and select 9 of them as training and 1 subset as testing in turn, so that each subset will be used as On the test data, the cross-validation accuracy is calculated based on the proportion of correctly classified samples. Change different parameter combinations, that is, the penalty coefficient C and the kernel parameter γ value in the kernel function, and select the combination that obtains the best cross-validation accuracy as the single-class support vector machine classification model parameters, and at the same time, w and b are determined accordingly. Using the determined classification model, single-class classification is carried out on the entire remote sensing image, and the water body class extraction result is obtained.

The single-class support vector machine only needs the samples of the river target category to learn, and can obtain the required classification results of the river target category, which reduces the tasks in the training stage. However, using multi-class classification methods to extract a single target category requires dividing the image into multiple object categories, which greatly increases the workload.

In this embodiment, Fig. 3 and Fig. 4 respectively show the river target extraction results obtained by using the nearest neighbor classification and Bayesian classification methods. Using the Bayesian classification method, it is necessary to divide the experimental images into four categories: impermeable surface, woodland, grassland and water body, and select training samples for each category, which increases the workload and uncertainty of the learning stage, making the overall Classification difficulty increases. Compared with the one-class SVM method using one class of training samples in Figure 2, the classification results are very similar, and the training of the one-class SVM method is easier.

D. Fine detection process: Segment the image of the candidate area of the river to generate shape features: Based on the results obtained in the rough screening of the C link, use the image threshold segmentation method to extract the target candidate area and generate shape features at the same time. Since there are only white and black pixels in the river candidate area image, set the threshold parameter Th=10, merge the adjacent pixels that are greater than the threshold, and generate target category connected areas of different sizes, and then calculate from the pixels in each connected area Shape features, the calculation method is as formula (3).

According to the actual river target inspection requirements, small area candidate targets are removed, and small area removal is carried out with the area threshold AT=50, that is, candidate targets with area feature A<50 are removed and merged into the background area. The result is shown in Figure 5.

According to the actual application requirements, the river target should have a certain area, and the relevant shape characteristics are described as follows.

Border length e: the number of border pixels, the border length of a pixel is 1.

Area A: The total number of pixels that make up the object, and the length of a pixel edge is set to 1.

Shape feature index:

The square root of the boundary length divided by 4 times the area, S describes the relationship between the object boundary and the area, the longer the boundary, the larger the value.

E. Determine the river target according to the shape feature index to realize target detection: shape information is a very important factor in the visual discrimination of remote sensing images. For a remote sensing image, the shape information is extracted by segmentation to obtain a multi-attribute A collection of image objects. In this collection, the shape features of each image object are matched with the attributes of river targets to filter out river targets. The shape feature of the river target mainly has a large shape feature index, and the value 2.5 is selected as the threshold for judgment.

On the basis of removing the small area target in the D step, according to the shape feature index of the candidate target generated by the segmentation, the threshold of the shape feature index is set to ST=2.5, and the river target feature matching is carried out. When the shape feature index value S>2.5 of the candidate target area is retained as the target matching result, if the shape feature index value S≤2.5 is removed as the background, the detection result is shown in Figure 6.

In the process of small area target removal and feature matching, the set administrative feature threshold should be adjusted according to actual needs, and different applications require different parameters.

The invention adopts a single-class support vector machine to detect the river target, has simple flow, greatly reduces calculation amount, accelerates work efficiency, and is worthy of popularization and use.

Claims (6)

1. A method utilizing a single-class support vector machine to detect river targets is characterized in that: comprising the following steps: A. Select the band: Select the band combination in the remote sensing image to distinguish the feature and extract the river, so as to identify the water body, vegetation and other ground objects; B. Feature analysis and selection: Select the image processed in step A, and perform spectral feature analysis: the spectral feature used in spectral analysis is: rivers and other water targets in remote sensing images, distinguish river targets by gray value; C. Rough screening process: based on spectral features, extract river candidate areas: select several target samples, extract the spectral value of each pixel as a classification feature, and use the formula x _i = (r _i , g _i , b _i ) to generate a feature vector, Use the generated feature vector to solve the optimization problem of the single-class support vector machine, and use the RBF kernel function during training Carry out training, and determine the classification model through 10-fold cross-validation, perform single-class classification on the entire remote sensing image, and obtain the water body category extraction result; Among them, r _i is the red component, g _i is the green component, b _i is the blue component; Among them, || _xi -x _j || represents the Euclidean distance between any two points x _i and x _j in the space; γ is the kernel parameter; D. Fine detection process: According to the results in the C link, image segmentation is performed on the candidate area of the river, and the shape feature index is generated: On the basis of the results obtained in the rough screening of the C link, the image threshold segmentation technology is used to set the threshold parameter, which will be greater than Threshold and adjacent pixels are merged to generate connected regions of target categories of different sizes, and then the shape feature index is calculated from the pixels in each connected region And set the area threshold to remove the small area, and finally merge the image with the removed small area into the background area; Among them, the border length e: the number of border pixels, the border length of a pixel is 1; Area A: the total number of pixels that make up the object, and the length of one pixel edge is set to 1; E. Target detection: set the shape index threshold, and determine the river target according to the shape feature index obtained in the D link, and realize the target detection. 2. the method for utilizing a single-class support vector machine to detect river targets according to claim 1 is characterized in that: the band selection link selects the 4, 3, and 2 bands of the remote sensing image, and assigns red, green, and blue respectively color to distinguish features. 3. The method for detecting river targets by using a single-class support vector machine according to claim 1, characterized in that: the gray value of the feature analysis and selection link is 10-20. 4. the method for utilizing single class support vector machine to detect river target according to claim 1, is characterized in that: the threshold parameter when image segmentation is in the described fine detection process is: Th=10; Area threshold is: AT= 50. 5. The method for detecting river targets by using a single-class support vector machine according to claim 1, characterized in that: the shape index threshold in the target detection link is ST=2.5. 6. the method utilizing single class support vector machine to detect river target according to claim 1 is characterized in that: the process of optimization problem solution in the described C link is: given 1 training samples, then have: And satisfy (w ^T φ(x _i )+b)≥1-ξ _i , ξ _i ≥0 Among them, C>0 is the penalty coefficient, φ( _xi ) is the function that maps the vector xi to the high-dimensional space; w is the normal vector of the classification surface in the feature space, w ^T is the transpose of w, b is the intercept of w distance, and _ξi is the relaxation factor.