WO2018076138A1

WO2018076138A1 - Target detection method and apparatus based on large-scale high-resolution hyper-spectral image

Info

Publication number: WO2018076138A1
Application number: PCT/CN2016/103070
Authority: WO
Inventors: 李岩山; 徐健杰; 黄庆华; 夏荣杰; 谢维信; 刘鹏
Original assignee: 深圳大学
Priority date: 2016-10-24
Filing date: 2016-10-24
Publication date: 2018-05-03

Abstract

A target detection method and apparatus based on a large-scale high-resolution hyper-spectral image. The method comprises: reading a hyper-spectral image corresponding to a target (S101); pre-processing the hyper-spectral image (S102); detecting all candidate spatial spectral domain interest points of the hyper-spectral image to obtain a first set (S103); screening all the candidate spatial spectral domain interest points in the first set according to a response intensity to obtain a second set (S104); performing spectral angle matching according to a spectral curve corresponding to the second set to obtain an image block of a potential target area (S105); performing feature description on the image block, and encoding same to obtain a vector corresponding to the image block (S106); calculating the value of a classification function corresponding to the image block according to the vector corresponding to the image block (S107); if the value of the classification function corresponding to the image block is greater than a classification threshold value, determining that the image block contains the target (S108); if the value of the classification function corresponding to the image block is less than or equal to the classification threshold value, splitting the image block (S109); splitting all the candidate spatial spectral domain interest points according to sub-image blocks to form a first set corresponding to the sub-image blocks (S110); and repeatedly operating on the split sub-image blocks sequentially until the value of a classification function corresponding to a certain sub-image block obtained by means of splitting is greater than the classification threshold value, or a sub-image block obtained by means of splitting reaches a specified minimum size (S111). The detection effect in target detection of a high-resolution hyper-spectral image can be improved.

Description

Target detection method and device based on large-scale high-resolution hyperspectral image

Technical field

The invention belongs to the field of information technology, and in particular relates to a target detection method and device based on large-scale high-resolution hyperspectral images.

Background technique

Hyperspectral imagery relative to grayscale, RGB (Red, Red; Green, Green; Blue, Blue) color maps, containing spatial and spectral information, large amounts of data, in detecting camouflage and concealed military targets, and in civilian search and rescue Aspects such as detection targets have an important role. With the development of hyperspectral technology, the current hyperspectral image exhibits high spatial resolution. The ground object has rich texture and structural information on the hyperspectral image, and the spectral information it contains is also very complicated and rich. The existing hyperspectral image target detection algorithm mainly uses spectral information for target detection, including Orthogonal Subspace Projection (OSP), Generalized Likelihood Ratio Test (GLRT), and constrained energy minimization. Algorithm (Constrained Energy Minimization, CEM), Adaptive Cosine Estimator (ACE), RX Anomaly Detection Algorithm (RXD), etc. These algorithms are generally applicable to satellite remote sensing hyperspectral images with low spatial resolution. The existing hyperspectral image target detection algorithm mainly utilizes the spectral information of the pixels, and their detection effect in the target detection of the high-resolution hyperspectral image is poor.

Summary of the invention

In view of this, embodiments of the present invention provide a target detection method and apparatus based on large-scale high-resolution hyperspectral images, so as to solve the problem that the prior art has poor detection effect in high-resolution hyperspectral image target detection. problem.

In a first aspect, an embodiment of the present invention provides a mesh based on a large-scale high-resolution hyperspectral image. Standard detection methods, including:

Reading a hyperspectral image corresponding to the target;

Preprocessing the hyperspectral image;

Detecting all candidate null spectral domain points of interest of the preprocessed hyperspectral image to obtain a first set;

Selecting candidate spatial domain points of interest in the first set according to response strength to obtain a second set;

Performing spectral angle matching according to the spectral curve corresponding to the second set to obtain an image block of a potential target area;

Characterizing the image block and encoding to obtain a vector corresponding to the image block;

Calculating a value of a classification function corresponding to the image block according to a vector corresponding to the image block;

If the value of the classification function corresponding to the image block is greater than a classification threshold, determining that the image block includes the target;

If the value of the classification function corresponding to the image block is less than or equal to the classification threshold, segmenting the image block until the value of the classification function corresponding to the segmented sub-image block is greater than the classification threshold, or The segmented sub-image block reaches the specified minimum size.

In a second aspect, an embodiment of the present invention provides a target detecting apparatus based on a large-scale high-resolution hyperspectral image, including:

a reading module for reading a hyperspectral image corresponding to the target;

a preprocessing module for preprocessing the hyperspectral image;

a detecting module, configured to detect all candidate spatial domain points of interest of the preprocessed hyperspectral image to obtain a first set;

a screening module, configured to filter candidate spatial domain points of interest in the first set according to response strength to obtain a second set;

a spectral angle matching module, configured to perform spectral angle matching according to the spectral curve corresponding to the second set to obtain an image block of a potential target area; and a feature description module configured to describe the image block, And encoding to obtain a vector corresponding to the image block;

a calculation module, configured to calculate a value of a classification function corresponding to the image block according to a vector corresponding to the image block;

a target determining module, configured to determine that the image block includes the target if a value of a classification function corresponding to the image block is greater than a classification threshold;

a segmentation module, configured to: if the value of the classification function corresponding to the image block is less than or equal to the classification threshold, segment the image block until a value of a classification function corresponding to the segmented sub-image block is greater than The classification threshold is described, or the segmented sub-image block reaches a specified minimum size.

Compared with the prior art, the embodiment of the present invention has the beneficial effects that the embodiment of the present invention separates the hyperspectral image by using the spatial spectral domain to extract the interest points, and separates the interest points of the optical spectrum domain, and describes, by description, The coding and classification realize the fast recognition and positioning of the target object, thereby improving the detection effect in the target detection of the high-resolution hyperspectral image.

DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments or the description of the prior art will be briefly described below. It is obvious that the drawings in the following description are only the present invention. For some embodiments, other drawings may be obtained from those of ordinary skill in the art in light of the inventive workability.

FIG. 1 is a flowchart showing an implementation of a target detection method based on a large-scale high-resolution hyperspectral image according to an embodiment of the present invention;

2 is a schematic diagram showing an original spectral curve of a point in a hyperspectral image and a sampling curve whose sampling frequency is 1/3 of the original spectral curve in the embodiment of the present invention;

FIG. 3 is a schematic diagram showing quadtree partitioning in an embodiment of the present invention; FIG.

FIG. 4 is a schematic diagram showing an overall framework adopted by a target detection method based on a large-scale high-resolution hyperspectral image according to an embodiment of the present invention; FIG.

5 to 8 illustrate the use of a large scale high resolution based on an embodiment of the present invention in different scenarios. A schematic diagram of the target detection method of the spectral image and the effect of using other algorithms;

FIG. 9 is a structural block diagram of a target detecting apparatus for a hyperspectral image according to an embodiment of the present invention.

detailed description

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

FIG. 1 is a flowchart showing an implementation of a target detection method based on a large-scale high-resolution hyperspectral image according to an embodiment of the present invention. As shown in Figure 1, the method includes:

In step S101, the hyperspectral image corresponding to the target is read.

The hyperspectral image may be a large-scale high-resolution hyperspectral image. For example, the hyperspectral image may have a size of M×N×B ₁ , where M represents the number of rows of the hyperspectral image and N represents the hyperspectral image. The number of columns, B ₁ represents the number of original bands of the hyperspectral image.

In step S102, the hyperspectral image is preprocessed.

In a possible implementation manner, preprocessing the hyperspectral image includes: performing band sampling processing on the hyperspectral image. In this implementation manner, in order to make the target detection method based on the large-scale high-resolution hyperspectral image have a certain real-time property, the original hyperspectral image can be subjected to band sampling processing. For example, the sampling interval may be k, the size of the hyper-spectral image band through the sampling process may be M × N × B _2, wherein, B ₂ represents a high band spectrum of the sampled image. B ₂ ={1,1+k,1+2k, . . . , B _max }, B _max ≤B ₁ , where B _max represents the maximum number of bands of B ₂ .

2 is a schematic diagram showing an original spectral curve of a point in a hyperspectral image and a sampling curve whose sampling frequency is 1/3 of the original spectral curve in the embodiment of the present invention. In Fig. 2, the horizontal axis is the wavelength and the vertical axis is the original digital value recorded by the sensor. It can be seen from Fig. 2 that in the sampled spectral curve, the spectral domain dimension is greatly reduced, and the amount of data is also greatly reduced. In the post-use spectral curve, the spectral information is preserved while reducing the amount of data, which provides a basis for the rapid detection of spatio-temporal points of interest.

In step S103, all candidate spatial domain points of interest of the preprocessed hyperspectral image are detected to obtain a first set.

In the embodiment of the present invention, the null spectral domain interest point may refer to a point where the pixel value of the entire hyperspectral cube changes drastically. In an embodiment of the invention, all candidate spatial domain points of interest of the hyperspectral image may be extracted prior to pre-processing. It is assumed that p _j represents the j-th candidate empty spectral domain interest point, v _j represents the response intensity corresponding to the j-th candidate empty spectral domain interest point, and p _j to v _j are a one-to-one mapping relationship, G:p _j →v _j . The first set P={p ₁ , p ₂ , p ₃ , . . . , p _m |G(p _x )≥G(p _x+1 )} can be constructed according to all candidate spatial domain points of interest, wherein P Representing the first set, x ∈ {1, 2, 3, ..., m-1}, m represents the total number of candidate spatial domain points of interest in the hyperspectral image.

In step S104, the candidate null spectral domain points of interest in the first set are filtered according to the response strength to obtain a second set.

Since the data volume of the candidate spatial domain interest points is very large, the candidate spatial domain interest points in the first set can be filtered to obtain a second set, and the second set is used to represent the hyperspectral image to improve the target. The efficiency of detection. For example, F _n can be used to form a new subset of the first n elements of the first set, and can be used.

Represents a subset of hyperspectral images,

among them,

Represents the second set, n < m. Usually, m is much larger than n.

In a possible implementation, the candidate empty spectral domain points of interest in the first set are filtered according to the response strength, and the second set is obtained, including: filtering the first n candidate spaces with the highest response strength from the first set. The spectral domain points of interest result in a second set, where n is a positive integer.

In step S105, spectral angle matching is performed according to the spectral curve corresponding to the second set to obtain an image block of the potential target area.

In a possible implementation manner, the method further includes: performing spectral angle matching on the spectral curve corresponding to the candidate empty spectral domain interest point by using the target corresponding spectral curve set to exclude the image block of the non-potential target area. For example, the spectral angle similarity threshold can be h. In this implementation, in order to improve the computational efficiency, the amount of subsequent calculations is reduced. Firstly, spectral peak matching can be performed by using the spectral curve set corresponding to the target and the spectral curve corresponding to the candidate spatial spectral domain, and the candidate of the sampled hyperspectral image can be calculated. The spectral angle similarity of the spectral curve set corresponding to the target in the spatial domain and the target, and then using the similarity to compare with the set spectral angle similarity threshold h, if the similarity is lower than the spectral angle similarity threshold h, It is determined that there is a possibility that the image block has a target, otherwise the image block is excluded. In step S106, the image block is characterized and encoded to obtain a vector corresponding to the image block.

An example of embodiment of the present invention may be employed 3D SIFT p _j to be described is formed _{q j, j∈ {1,2,3, ...} , m}, the mapping between the p _j q _j can be used with H is expressed as H:p _j →q _j . For the second set

Describe and form a description set

The description set obtained by each image cube is subjected to a word bag model (BoW) coding to form a statistical distribution histogram, and a vector corresponding to each tile image can be obtained.

In step S107, the value of the classification function corresponding to the image block is calculated from the vector corresponding to the image block.

As an example of the embodiment of the present invention, after obtaining the vector corresponding to each image block, the hyperspectral image may be represented by the obtained vector, and may be classified by using a SVM (Support Vector Machine) classifier. Assuming that the vector corresponding to an image block is X, the classification function of the discrimination target may be f(X)=w ^T X+b, where the w vector and the b vector are obtained after the SVM classifier trains a large number of target samples. vector.

In step S108, if the value of the classification function corresponding to the image block is greater than the classification threshold, it is determined that the image block contains the target.

For example, the classification threshold can be Td. If f(X)≥Td, it is determined that the image block corresponding to the vector X contains the target, otherwise the current image is segmented by the quadtree partitioning method.

In step S109, if the value of the classification function corresponding to the image block is less than or equal to the classification threshold, the image block is segmented.

In a possible implementation manner, segmenting an image block includes: performing quadtree partitioning on the image block. It should be noted that in this implementation, the quadtree partitioning of the hyperspectral image is performed on the airspace. Different from the general quadtree partitioning, since the partitioning block size in the image spatial domain direction has little influence on the extraction of the interest points, it is not necessary to fill the image so that the rows and columns are all 2 nth power. For example, suppose the size of the hyperspectral image is M × N × B, where M is the number of rows, N is the number of columns, B For the number of bands. When splitting, according to the M and N parity, do the following split:

Where R _i is a split sub-cube, the goal of the quad-tree segmentation method is to make the hyperspectral image block in the large environment contain fewer objects, and the class of the class after the feature mapping is higher.

In step S110, all candidate spatial domain points of interest are segmented according to the sub-image block to form a first set of corresponding sub-image blocks.

In a possible implementation manner, after performing quadtree partitioning on the hyperspectral image, P _i is a candidate point set of the R _i region, expressed as P _i ={p ₁ , p ₂ , p ₃ ,..., p _yi |G(p _z )≥G(p _z+1 )}, where yi is the maximum value of the number of candidate spatial spectral points of interest corresponding to the subcube R _i , z ∈ {1, 2, 3, ..., yi-1}. Simultaneously,

Also forming four subsets

Expressed as

Continuing to select n candidate spatial domain points of interest in the segmented sub-cube to form a set P′ _i , the set P′ _{i is} used to represent a hyperspectral image, and P′ _i can be expressed as P′ _i =F _n (P _i ), the feature description set Q _{i of} the corresponding image block can be expressed as

The newly formed image block feature description is concentrated, and it is not necessary to describe the n candidate empty spectral domain points of interest, but only operations

Part of it,

Then inherit from the parent image.

In step S111, the operations of the divided sub-images are sequentially repeated until the value of the classification function corresponding to the segmented sub-image block is greater than the classification threshold, or the divided sub-image block reaches the specified minimum size.

In a possible implementation, after calculating the value of the classification function corresponding to the image block according to the vector corresponding to the image block, the method further includes: if the value of the classification function corresponding to the sub-image block is less than or equal to the classification threshold, proceeding to The sub-image block is subjected to quadtree partitioning until a sub-image obtained by the segmentation The value of the classification function corresponding to the block is greater than the classification threshold or the specified minimum size is reached to stop the segmentation.

The data of the original hyperspectral image is taken as the root node, and is successively recursively divided into four sub-cubes until the condition f(X)≥Td is satisfied. f(X) is the classification function of the target map corresponding to the target map of the hyperspectral image after the interest point and feature description based on the spatial spectral domain extraction, and Td is the classification threshold. It should be noted that since the selected candidate spatial spectral domain interest points may contain other targets, the value of f(X) may be affected. If only one target object is included, after calculation, the obtained target object's f ( X) is larger.

FIG. 3 shows a schematic diagram of quadtree partitioning in an embodiment of the present invention. As shown in FIG. 3, it is assumed that the original hyperspectral image is input with a sequence number of 0, and the sub-cube blocks that are sequentially divided are encoded. The longer the serial number length, the deeper the depth of the tree, and the smaller the size of the sub-image block.

It should be noted that the general target object will be composed of smaller image blocks, and the image determined by the classification function may be misclassified, and a certain post-processing is required. For example, by finding the centroid of the image block determined to be the target, the centroid of the target can be calculated. Suppose the target centroid is C, the centroid of each image block _C α, then

Where β is the number of image blocks after the target is determined, and after the centroid is obtained, the position of the target object can be determined, so that the graphic block with the size and shape set in advance can be used for positioning.

FIG. 4 is a schematic diagram showing an overall framework adopted by a target detection method based on a large-scale high-resolution hyperspectral image according to an embodiment of the present invention. As shown in FIG. 4, as an example of the embodiment of the present invention, candidate spatial domain points of interest may be first extracted in a cube corresponding to the hyperspectral image, and the non-potential target region is quickly excluded according to the spectral curve corresponding to the candidate spatial domain of interest points. Characterizing the suspected target area, then using BoW for image coding, and classifying it with the SVM classifier. If it is judged as the target, the block target detection ends, otherwise the image is divided by the quadtree segmentation method. Repeat the above steps and iterate until the condition is met.

It should be understood that, in the embodiment of the present invention, the size of the sequence numbers of the foregoing processes does not mean that the execution is smooth. The order of execution of the processes should be determined by their functions and internal logic, and should not be construed as limiting the implementation process of the embodiments of the present invention.

In the embodiment of the present invention, the hyperspectral image is segmented by using the spatial spectral domain, the hyperspectral image is segmented, the interest points of the spatial domain are separated, and the target is quickly identified and positioned by description, coding, and classification, thereby improving high resolution. The detection effect of the target in the detection of the hyperspectral image.

In order to more intuitively show the effect of the embodiment of the present invention in target detection based on large-scale high-resolution hyperspectral images, experimental results are shown below. The hyperspectral image used in this experiment is from the hyperspectral camera model Image-λ-V10E-HR, ZOLIX INSTRUMENTS CO., LTD., with 728 bands, wavelength range from 362.05 to 1002.47 nm, spatial resolution of 2.9 cm/pixel. The spectral resolution reached 0.88 nm. In this experiment, taking the detection of a car as an example, the n value of F _n is set to 100, the classification threshold Td is 1.1, the spectral angle similarity threshold h is 0.19, and the sampling interval k is 3. Experiments were carried out under different scenarios, and the target detection methods based on large-scale high-resolution hyperspectral images provided by the present embodiment were compared with the conventional RXD, CEM, ACE and the like algorithms. Figures 5 to 8 set different scenes, in which (a) is the original hyperspectral image, (b) is the experimental result of the embodiment, and the black dots on the image in (c) to (f) are Target (c) is the result of binarizing the RXD algorithm using the threshold 300, and (d) is the result of binarizing the CEM algorithm using the threshold of 0.2, and (e) is using the threshold of 0.08. The results of the binarized ACE algorithm of the detection result, Fig. (f) is the result of binarizing the MF algorithm by using the threshold value of 0.2; the RXD detection does not require prior knowledge, and the CEM, ACE and MF detections in Fig. 5 to Fig. 7 are The spectral curve of a certain point on the front cover of the car is taken as a priori knowledge, while in Figure 8, the spectral curve at a certain point on the roof of the car is taken as a priori knowledge.

As shown in Fig. 5, there are white cars, trees, roads and street lamps in the scene 1. With this embodiment, the position of the car can be accurately detected. In addition to detecting the car, the RXD algorithm, CEM and MF are also recognized as Target, and ACE was not accurately detected.

As shown in Fig. 6, in the second scene, there are black cars, roads, sidewalks and soils. With this embodiment, the position of the car can be better identified. For black cars, RXD, CEM, and MF can't detect cars, and the target of ACE detection is concentrated in the black car position, but the effect is poor.

As shown in Fig. 7, in the third scene, there are white cars, roads, sidewalks and soils. The biggest difference compared to the scene is that the background is replaced, but the car can be better detected by this embodiment. The detection result is similar to that of scene 1. The RXD algorithm detects other non-target substances, and the CEM algorithm and MF algorithm can roughly detect the target, while the ACE algorithm is less effective.

As shown in Figure 8, there are three cars of different colors, roads and vegetation in scene four. The results of RXD algorithm detection have other non-automobile targets in addition to the car. The results of CEM algorithm and MF algorithm to detect the upper vehicle are more obvious. Two cars can also be detected, but the effect is less obvious. The ACE algorithm can only detect the upper vehicle, and the other two vehicles cannot be detected clearly.

According to the principle of RXD, when there is an abnormal target, the corresponding energy will be relatively small, and the calculated result will be a relatively large value. When the difference between the target and the background is large, the detection result is ideal, and the biggest advantage of the algorithm is that no target is needed. Prior Knowledge. ACE needs to know the sample covariance matrix. It is necessary to estimate the covariance matrix of all target sample pixels. When the target pixel has less hyperspectral image, the impact on the estimated value is small, but the target of this experiment is large, which affects Detection effect, ACE will be better for a small number of target detection. Both CEM and MF require prior knowledge of the target. By suppressing the background detection target, the two methods are also most effective for small target detection.

The embodiments of the present invention are well suited for use in large-scale high-resolution hyperspectral image scenarios, because the method of detecting feature points combines spatial and spectral information to maximize the amount of information provided by hyperspectral images. The SVM model of the target training object in advance, the feature description is divided by the quadtree method, and the nature of the image block is distinguished according to the description, so that the target object can be accurately detected.

FIG. 9 is a structural block diagram of a target detecting apparatus for a hyperspectral image according to an embodiment of the present invention. For convenience of explanation, only parts related to the embodiment of the present invention are shown in FIG.

As shown in FIG. 9, the device includes: a reading module 90 for reading a hyperspectral image corresponding to a target; a preprocessing module 91 for preprocessing the hyperspectral image; and a detecting module 92 for detecting The pre-processed all the candidate spatial domain points of interest of the hyperspectral image are obtained, and the screening module 93 is configured to filter the candidate spatial domain points of interest in the first set according to the response strength, a second set; a spectral angle matching module 94, configured to enter, according to the spectral curve corresponding to the second set Performing a spectral angle matching to obtain an image block of a potential target area; a feature description module 95 for characterizing the image block and encoding a vector corresponding to the image block; and a calculation module 96 for determining an image according to the image a vector corresponding to the block is used to calculate a value of the classification function corresponding to the image block; and a target determining module 97 is configured to determine that the image block includes the target if the value of the classification function corresponding to the image block is greater than a classification threshold; The module 98 is configured to: if the value of the classification function corresponding to the image block is less than or equal to the classification threshold, segment the image block until the value of the classification function corresponding to the segmented sub-image block is greater than The classification threshold is described, or the segmented sub-image block reaches a specified minimum size.

In a possible implementation, the screening module 93 is specifically configured to: screen, from the first set, the first n candidate spatial domain points of interest with the highest response strength, to obtain a second set, where n is A positive integer.

In a possible implementation, the device further includes: the spectral angle matching module 94 is further configured to: perform spectral angle matching by using a spectral curve corresponding to the target spatial spectral interest point by using the spectral curve set corresponding to the target, Exclude image blocks from non-potential target areas.

In a possible implementation, the segmentation module 98 is specifically configured to perform quadtree partitioning on the image block.

Those of ordinary skill in the art will appreciate that the modules and algorithm steps of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the solution. A person skilled in the art can use different methods for implementing the described functions for each particular application, but such implementation should not be considered to be beyond the scope of the present invention.

A person skilled in the art can clearly understand that for the convenience and brevity of the description, the specific working process of the foregoing device and module can refer to the corresponding process in the foregoing method embodiment, where No longer.

In the several embodiments provided by the present application, it should be understood that the disclosed apparatus and method may be implemented in other manners. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there may be another division manner. For example, multiple modules may be combined or integrated. Go to another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interface, indirect coupling or communication connection of the module, and may be in electrical, mechanical or other form.

The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

The functions may be stored in a computer readable storage medium if implemented in the form of a software functional unit and sold or used as a standalone product. Based on such understanding, the technical solution of the present invention, which is essential or contributes to the prior art, or a part of the technical solution, may be embodied in the form of a software product, which is stored in a storage medium, including The instructions are used to cause a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and the like. .

The above description is only a specific embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily think of within the technical scope disclosed by the present invention. Variations or substitutions are intended to be covered by the scope of the invention. Therefore, the scope of the invention should be determined by the scope of the claims.

Claims

A target detection method based on large-scale high-resolution hyperspectral image, characterized in that it comprises:

Reading a hyperspectral image corresponding to the target;

Preprocessing the hyperspectral image;

Detecting all candidate null spectral domain points of interest of the preprocessed hyperspectral image to obtain a first set;

Selecting candidate spatial domain points of interest in the first set according to response strength to obtain a second set;

Performing spectral angle matching according to the spectral curve corresponding to the second set to obtain an image block of a potential target area;

Characterizing the image block and encoding to obtain a vector corresponding to the image block;

Calculating a value of a classification function corresponding to the image block according to a vector corresponding to the image block;

If the value of the classification function corresponding to the image block is greater than a classification threshold, determining that the image block includes the target;

If the value of the classification function corresponding to the image block is less than or equal to the classification threshold, segmenting the image block until the value of the classification function corresponding to the segmented sub-image block is greater than the classification threshold, or The segmented sub-image block reaches the specified minimum size.
The method according to claim 1, wherein the candidate empty spectral domain points of interest in the first set are filtered according to the response strength to obtain a second set, comprising:

Selecting the first n candidate null spectral domain points of interest with the highest response strength from the first set, and obtaining a second set, where n is a positive integer.
The method of claim 1 further comprising:

Spectral angle matching is performed by using a spectral curve set corresponding to the target and a spectral curve corresponding to the candidate empty spectral domain interest point to exclude image blocks of the non-potential target area.
The method according to claim 1, wherein the segmenting the image block comprises: performing quadtree partitioning on the image block.
A target detecting device based on a large-scale high-resolution hyperspectral image, comprising:

a reading module for reading a hyperspectral image corresponding to the target;

a preprocessing module for preprocessing the hyperspectral image;

a detecting module, configured to detect all candidate spatial domain points of interest of the preprocessed hyperspectral image to obtain a first set;

a screening module, configured to filter candidate spatial domain points of interest in the first set according to response strength to obtain a second set;

a spectral angle matching module, configured to perform spectral angle matching according to the spectral curve corresponding to the second set to obtain an image block of a potential target area; a feature description module, configured to describe the image block, and encode the The vector corresponding to the image block;

a calculation module, configured to calculate a value of a classification function corresponding to the image block according to a vector corresponding to the image block;

a target determining module, configured to determine that the image block includes the target if a value of a classification function corresponding to the image block is greater than a classification threshold;

a segmentation module, configured to: if the value of the classification function corresponding to the image block is less than or equal to the classification threshold, segment the image block until a value of a classification function corresponding to the segmented sub-image block is greater than The classification threshold is described, or the segmented sub-image block reaches a specified minimum size.
The apparatus according to claim 5, wherein the screening module is specifically configured to:

Selecting the first n candidate null spectral domain points of interest with the highest response strength from the first set, and obtaining a second set, where n is a positive integer.
The device according to claim 5, wherein the spectral angle matching module is further configured to:

Using the spectral curve set corresponding to the target and the spectral curve corresponding to the candidate empty spectral domain interest point Spectral angle matching is performed to exclude image blocks from non-potential target areas.
The apparatus according to claim 5, wherein said segmentation module is configured to:

The image block is subjected to quadtree partitioning.