CN105005789B - A kind of remote sensing images terrain classification method of view-based access control model vocabulary - Google Patents

Abstract

The invention discloses a visual vocabulary-based classification method for remote sensing images, comprising the following steps: first, all remote sensing images are divided into a training set and a test set, and each remote sensing image is cut with a fixed size to obtain a preliminary slice map, Extract the preliminary slice image containing the target; generate a multi-layer Gaussian space pyramid through Gaussian blur and sampling for the preliminary slice image containing the target; perform SIFT feature extraction and LBP feature extraction on the image of each layer; The remote sensing words of the remote sensing image are clustered to obtain multiple cluster centers, and all the cluster centers form a remote sensing dictionary; different radius values are set, and frequency histograms are established for the remote sensing words within different radius values of each preliminary slice image: The frequency histogram of the remote sensing words in the training set is trained using the support vector machine RBF-SVM, and then the trained RBF-SVM is used to classify the remote sensing images in the test set.

Description

A Classification Method of Remote Sensing Image Objects Based on Visual Lexicon

technical field

The invention belongs to the technical field of remote sensing image processing, and relates to a ground object classification method based on high-resolution remote sensing images.

Background technique

With the rapid development of remote sensing satellites, the resolution of obtained remote sensing satellite images is getting higher and higher, and the resulting processing requirements for remote sensing images are also getting higher and higher. Among them, the object classification of high-resolution remote sensing satellites Identification is an important research direction, especially the scientificity of urban configuration, the distribution planning of agricultural planting and the classification and extraction of military sensitive targets are all of great significance. At the same time, the high resolution of high-resolution remote sensing satellites also means that the details are richer, and the background has a greater influence on the target, which brings challenges to image classification.

In the field of remote sensing image classification, there are mainly two types of classification methods: one is unsupervised method, which is realized by using the idea of clustering, including k-means, isodata and other methods. However, as the resolution of remote sensing images becomes higher and higher, the variance between classes will decrease, and the variance within classes will increase, resulting in overlapping between classes, the phenomenon of the same object with different spectra and the same spectrum with different objects, resulting in classification errors; the other type is The supervised method is also a commonly used classification method now, using the idea of learning and training, including BP neural network model, genetic model and support vector machine, etc. The supervised method has a certain incremental adjustment, but when the remote sensing image resolution When it keeps increasing, it is a challenge for the selection of training samples and the training speed of each type of samples, especially the selection of training samples is a process of manual selection, which brings greater time cost.

At the same time, the development of high-resolution remote sensing images means that the types of feature selection are increasing, and the relationship between the low-level features such as color, texture, and shape and the high-level information of the sample will have the disadvantages of poor generalization and low applicability. In order to overcome the gap between low-level features and high-level semantic information of samples, a derivation model based on intermediate features of samples and high-level semantic information of samples was born. The BOV (Bag of Visual Words) model, as a representative of intermediate features, has made some progress in OBIA (object-based remote-sensing image analysis). However, the traditional BOV model method has its own defects, and cannot make good use of all the information of the image, such as spatial hierarchical information, etc., and the BOV model classification is mainly used in the classification and recognition of natural objects, and it is not well utilized for the classification of remote sensing images.

How to use the idea of bag of words to solve the classification of remote sensing images has become one of the important research directions in the field of remote sensing images. However, due to the variety of ground objects in remote sensing images, it is difficult to represent all types of remote sensing images with a single bag of feature words method and a single frequency histogram of remote sensing words.

Contents of the invention

In view of this, the present invention provides a visual vocabulary-based remote sensing image feature classification method, which solves the problem of insufficient classification accuracy of remote sensing image feature.

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

Step 1. Divide all remote sensing images into training set and test set. For each remote sensing image, cut it with a fixed size to obtain a preliminary slice map, and extract the preliminary slice map containing the target.

Step 2: Generate a multi-layer Gaussian space pyramid through Gaussian blurring and sampling for the preliminary slice image containing the target extracted in step 1.

Step 3, perform feature extraction on the image of each layer in the Gaussian space pyramid, feature extraction includes local features, namely SIFT feature extraction, and image texture features, namely LBP feature extraction.

When performing SIFT feature extraction, only one layer of Gaussian filtering is performed, and then the image is divided into 16×16 subdivision slices by means of a sliding window, and the SIFT feature vector is extracted according to the gradient change of each subdivision slice to obtain about The 128-dimensional SIFT feature vector of the subdivision slice map; in each subdivision slice map, determine the LBP feature vector about the subdivision slice map according to the set radius and the number of sampling points.

Then the SIFT feature vector and the LBP feature vector corresponding to each subdivision slice map are combined to form a remote sensing word; thereby generating a remote sensing word about the remote sensing image.

Among them, the Gaussian spatial pyramid at the highest level satisfies that the number of remote sensing words corresponding to its preliminary slice map is at least 1/2 of the number of remote sensing dictionaries.

For the preliminary slice image in each layer of Gaussian pyramid, the following steps 4-5 are used for processing:

Step 4, clustering the remote sensing words of all the remote sensing images in the training set and the test set to obtain multiple cluster centers, and all the cluster centers form a remote sensing dictionary.

In the clustering process, PCA principal component analysis method is used to reduce the dimension of remote sensing words, retain the dimension that contributes the most to the covariance, and set the data loss rate threshold disratio so that the data loss rate does not exceed disratio.

Step five, set different radius values, and establish a frequency histogram for the remote sensing words in different radius values of each preliminary slice map:

The abscissa of the histogram is each remote sensing word, and the ordinate is the frequency of remote sensing words appearing in the preliminary slice map within the current radius value. The Euclidean distance between two remote sensing words in the preliminary slice image, if the Euclidean distance between the current remote sensing word A and another remote sensing word B is the minimum value of the Euclidean distance between all other remote sensing words and B, then the frequency value of A increase by 1.

Step 6: Use the support vector machine RBF-SVM to train the frequency histogram of the remote sensing words in the training set, and then use the trained RBF-SVM to classify the remote sensing images in the test set.

If the number of layers of the Gaussian pyramid to which the remote sensing word belongs is at the bottom, the weight of the classifier in the corresponding support vector machine is assigned to be the largest, and the higher the number of layers, the smaller the weight of the corresponding classifier.

Further, when the image is divided into 16*16 subdivision slices by means of a sliding window, the step size of the sliding window is set between 1 and 8.

Further, the kernel function in the RBF-SVM support vector machine selects histogram intersectionkernel.

Further, in step 2, the K-means clustering method is used for clustering.

Beneficial effect:

1. This method first divides the remote sensing image set into two parts, the test set and the training set, and then extracts feature points from all remote sensing image slices to obtain the words of each slice, and then clusters the feature points of the training set to obtain a dictionary , and finally build a frequency histogram for the words in the dictionary and the test set for training, and use the word histogram of the test set for testing, which can more accurately classify remote sensing images and features, and the classification results are more accurate.

2. In this method, a multi-layer Gaussian pyramid method is used when performing Gaussian blur and sampling on the sliced image, so the spatial information of the remote sensing image can be better utilized, thereby making the classification more accurate.

3. In the traditional method, because the SIFT feature vector only contains the direction and gradient information in the 16*16 slice map, that is, the information of the stable point in the slice map, it does not have the texture feature information of the entire slice map. Therefore, based on the SIFT feature The rotation-invariant uniform LBP feature is added to it, which makes up for the texture feature information that does not exist in the SIFT feature, and can express all the information in the remote sensing image in an all-round way.

4. In this method, when generating the frequency histogram, select a plurality of different radii, and ensure that the selected radius should make the interior contain enough remote sensing words, and the remote sensing words and The frequency histogram generated by the remote sensing dictionary can meet the needs of generating the frequency histogram of remote sensing words, and at the same time use the spatial information from the center to the surrounding in the remote sensing image, so that the target in the remote sensing image can be better expressed and improved. classification accuracy.

5. This method finally uses a multi-classifier plus a decision maker to obtain the decision result, classifies the frequency histogram of remote sensing words at each layer in the Gaussian empty pyramid, and assigns weights to different classifiers. We believe that the information weight at the bottom of the Gaussian pyramid The value is the largest, and the higher the information weight is, the smaller the information weight is. After decision-making, the category of the remote sensing slice map is obtained. By assigning weights to each Gaussian pyramid classification result, not simply trusting the classification results of a certain layer of images, through different levels of Gaussian pyramids. The classification results jointly determine the target in the remote sensing image and improve the classification accuracy.

Description of drawings

Fig. 1 is the implementation flowchart of the present invention.

Fig. 2 is the generation flowchart of the remote sensing word of the present invention.

Fig. 3 is a flow chart for generating a remote sensing word frequency histogram in the present invention.

Fig. 4 is the design of the classifier of the present invention.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and examples.

Step 1. Divide all remote sensing images into a training set and a test set. For each remote sensing image, cut it with a fixed size to obtain a preliminary slice map, and extract the preliminary slice map containing the target;

Step 2. Generate a multi-layer Gaussian space pyramid through Gaussian blurring and sampling for the preliminary slice image containing the target extracted in step 1. The number of layers of the Gaussian space pyramid is determined according to the size of the preliminary slice image;

Accompanying drawing 2 is the flow chart of generation of remote sensing words, including two parts of Gaussian space pyramid generation and multi-feature extraction.

Firstly, multi-layer Gaussian space pyramids are generated by Gaussian blur and sampling for remote sensing slices. For slices of different sizes, Gaussian space pyramids with different layers can be generated. The pyramid at the highest level can meet at least 1/1 of the number of remote sensing dictionaries. 2. To form the remote sensing word frequency histogram, in this embodiment, there are 16 groups of eigenvalues, which can ensure that the generated frequency histogram is meaningful.

The multi-layer Gaussian spatial pyramid can make better use of the spatial information of remote sensing images, thus making the classification more accurate.

Secondly, multi-feature (SIFT feature and LBP feature) extraction is performed for each image in the Gaussian space pyramid. In order to maintain the consistency of the number of frequency histograms of each image, the SIFT feature is changed, only one layer of Gaussian filtering is performed, and the image is divided into subdivided slices by means of a sliding window. In this embodiment, it is a slice of 16*16 In the figure, the step size of the sliding window can be set from 1 to 8 according to actual needs, and a 128-dimensional SIFT feature vector is obtained according to the gradient change of each 16*16 subdivision slice map. In this way, the number of SIFT feature points for each image is consistent, which is convenient for subsequent histogram statistics.

Since the SIFT feature vector only contains the direction and gradient information in the 16*16 slice map, that is, the information of the stable point in the slice map, it does not have the texture feature information of the entire slice map, so the rotation invariant uniform LBP is added to the SIFT feature Features can make up for the representation of images, and can express all the information in remote sensing images in an all-round way. In each 16*16 slice image, the LBP feature is determined according to the radius and the number of sampling points (manually set). Since the rotation-invariant LBP feature is used, the dimensionality representing the image texture feature is less.

Combining the SIFT feature of each image with the LBP feature of uniform rotation is the remote sensing word representing this image (each 16*16 slice map corresponds to a remote sensing word), and the number of remote sensing words generated from each slice map is consistent.

Step 2: Clustering is performed based on all remote sensing words in the training set and the test set to obtain multiple cluster centers, and all the cluster centers form a remote sensing dictionary.

Since clustering is an offline process, with the increase of clustering data, the general sense of the overall division clustering method will produce geometric multiples of the calculation cost. Therefore, the idea of dimensionality reduction clustering is used to deal with large-scale and ultra-large-scale remote sensing dictionary data.

Here, the PCA principal component analysis method is used to reduce the dimensionality of the remote sensing words, and the dimension that contributes the most to the covariance is retained. However, it should be noted that the number of dimensionality reduction needs to ensure that a large amount of useful data is not lost. Therefore, the definition of the data loss rate disratio , ensuring that the rest of the data retains most of the useful information.

The clustering method adopts the K-means clustering method with balanced time efficiency and space efficiency. K-means clustering can directly obtain the cluster center, and the cluster center is the average value of all attributes in the class, which can well represent the class inner center. The number of cluster centers needs to be determined according to the number of classifications. The more categories that need to be classified, the more cluster centers are needed, which means that the more words in the remote sensing dictionary can meet the multi-category goals. classification accuracy requirements.

Step 3, set different radius values, and establish a frequency histogram for the remote sensing words in different radius values of each preliminary slice map:

The abscissa of the histogram is each remote sensing word, and the ordinate is the frequency of remote sensing words appearing in the preliminary slice map within the current radius value. The Euclidean distance between two remote sensing words in the preliminary slice image, if the Euclidean distance between the current remote sensing word A and another remote sensing word B is the minimum value of the Euclidean distance between all other remote sensing words and B, then the frequency value of A increase by 1.

Accompanying drawing 3 is the generation flow diagram of remote sensing word frequency histogram of the present invention, here in order to effectively utilize the space position information in the picture, for the space pyramid of different levels, choose suitable radius, to each layer slice figure in the pyramid is different The remote sensing words within the radius and the remote sensing dictionary generate a frequency histogram, and combine all the remote sensing word frequency histograms on this slice to represent the remote sensing word frequency histogram of this preliminary slice.

When selecting an appropriate radius, it should satisfy the requirement that it contains enough remote sensing words to meet the requirement of generating the frequency histogram of remote sensing words. The advantage of choosing multiple radii is to use the spatial information from the center to the surrounding in the remote sensing image, so that the target in the remote sensing image can be better expressed and the classification accuracy can be improved.

Step 4: Use the RBF-SVM support vector machine for training on the remote sensing word frequency histogram of the training set and the remote sensing frequency histogram of the test set, and select the histogram intersection kernel with good performance in the statistical field as a kernel function.

Attached Figure 4 shows the composition of the classifier. The decision result is obtained by using multiple classifiers plus a decision maker. The frequency histogram of remote sensing words in each layer of the Gaussian empty pyramid is classified, and the weights are assigned to different classifiers. We think that the Gaussian pyramid The information weight at the bottom layer is the largest, and the information weight is smaller as you go up, and the category of the remote sensing slice map is obtained through decision-making. By assigning weights to each Gaussian pyramid classification result, we do not simply trust the classification results of a certain layer of images, but use the classification results of different levels of Gaussian pyramids to jointly determine the target in the remote sensing image and improve the classification accuracy.

Since then, the object classification of remote sensing images has been completed.

To sum up, the above are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (4)

1. a kind of remote sensing images terrain classification method of view-based access control model vocabulary, it is characterised in that comprise the following steps：

Step 1: all remote sensing images are divided into training set and test set, for every width remote sensing images, with fixed size pair It, which cuts, obtains preliminary slice map, and extraction includes the preliminary slice map of target；

Step 2: more floor heights are generated by Gaussian Blur and sampling for the preliminary slice map comprising target extracted in step 1 This spatial pyramid；

Step 3: feature extraction, the feature extraction bag are carried out to each layer of image in the Gaussian spatial pyramid Include local feature i.e. SIFT feature extraction and the i.e. LBP feature extractions of image texture characteristic；

When carrying out SIFT feature extraction, one layer of gaussian filtering is only carried out, image is then divided into 16 by way of sliding window × 16 subdivision slice map, SIFT feature vector extraction is carried out according to the graded of each subdivision slice map and obtained on this Segment the SIFT feature vector of 128 dimensions of slice map；Radius and sampling number according to setting in each subdivision slice map is true The fixed LBP characteristic vectors on the subdivision slice map；

Then each SIFT feature vector sum LBP combination of eigenvectors corresponding to subdivision slice map forms a remote sensing word；Thus Generate the remote sensing word on the remote sensing images；

Wherein top Gaussian spatial pyramid meets correspondingly to obtain remote sensing word number in its preliminary slice map at least distant Feel the 1/2 of the quantity of dictionary；

The mode of following steps four~five is taken to be handled for the preliminary slice map in each layer of gaussian pyramid：

Step 4: being clustered to the remote sensing word of all remote sensing images in training set and test set, multiple cluster centres are obtained, All cluster centre composition remote sensing dictionaries；

In cluster process, dimensionality reduction is carried out using PCA PCAs to remote sensing word, retained to covariance contribution maximum Dimension, and set data loss rate threshold value disratio so that data loss rate is no more than disratio；

Step 5: setting different radii value, it is straight to establish frequency to the remote sensing word in each preliminary slice map different radii value Fang Tu：

The abscissa of the histogram is each remote sensing word, and ordinate is preliminary slice map of the remote sensing word in current radius value The frequency of middle appearance, the computational methods of the frequency are：Frequency values are initially 0, calculate in the preliminary slice map in current radius value Euclidean distance between remote sensing word two-by-two, if the Euclidean distance between current remote sensing word A to another remote sensing word B is all Frequency values increase by 1 of the remote sensing word to the minimum value, then A of B Euclidean distance；

Step 6: it is trained for the frequency histogram of remote sensing word in training set using SVMs RBF-SVM, then Terrain classification is carried out to remote sensing images in test set using the RBF-SVM after training；

If the number of plies of the gaussian pyramid belonging to remote sensing word in the bottom, distributes grader in its corresponding SVMs Maximum weight, more up, the weights of corresponding grader are smaller for the number of plies.

2. the remote sensing images terrain classification method of a kind of view-based access control model vocabulary as claimed in claim 1, it is characterised in that described When image to be divided into 16*16 subdivision slice map by way of sliding window, the step-length of sliding window is arranged between 1 to 8.

3. the remote sensing images terrain classification method of a kind of view-based access control model vocabulary as claimed in claim 1, it is characterised in that described Selection of kernel function histogram intersection kernel in RBF-SVM SVMs.

A kind of 4. remote sensing images terrain classification method of view-based access control model vocabulary as claimed in claim 1, it is characterised in that step K-means clustering methods are used in two, during cluster.