Hyperspectral fast target detection method based on compressed spectral imaging

Abstract

The present invention belongs to the technical field of hyperspectral target detection, and specifically relates to a hyperspectral rapid target detection method based on compressed spectral imaging. The present invention provides a hyperspectral rapid target detection method based on compressed spectral imaging, which utilizes the spectral coding template in the compressed spectral imaging technology to directly act the CEM operator on the radiation light of the spatial pixels entering the system, so that the system directly outputs the target detection measurement value of each spatial pixel. The present invention does not need to reconstruct the data cube, that is, the target detection research object is converted from three-dimensional spectral data to the radiation light of the target space, and the target detection result is directly output while performing push scanning, avoiding the problem of large data volume. There is almost no time cost in the coding template generation process. At the same time, the modulation speed of the spatial light modulator is very fast. Even if the target space is pushed and scanned, the spatial target detection can be quickly completed, providing a new solution for hyperspectral rapid target detection.

Description

[0001] Fields:

[0002] The present invention belongs to the technical field of hyperspectral target detection, and in particular relates to a hyperspectral rapid target detection method based on compressed spectral imaging.

[0003] Background:

[0004] Hyperspectral target detection is one of the important directions of hyperspectral remote sensing applications, involving fields including but not limited to environmental monitoring, mineral exploration, military reconnaissance, smart agriculture, etc. Compared with traditional target detection methods based on high spatial resolution remote sensing images, due to the unified characteristics of hyperspectral images, hyperspectral target detection mainly detects targets based on the difference in spectral characteristics between targets and background objects. This provides possibilities for target detection applications in situations where target size is limited, background objects are complexly distributed, or detector resolution is limited. Therefore, hyperspectral remote sensing has great advantages in target detection.

[0005] Domestic and foreign scholars have proposed many hyperspectral target detection algorithms to meet the needs of target detection, including classic algorithms such as the minimum distance method, spectral angle mapping, constrained energy minimization, and adaptive matched filtering. These algorithms are all based on target detection after hyperspectral imaging is acquired and reconstructed. However, hyperspectral imaging obtains rich data information by simultaneously detecting the two-dimensional geometric information and one-dimensional spectral information of the target space. Therefore, the reconstructed data is a three-dimensional spectral data cube, and due to the high spectral resolution, its data volume is often tens to hundreds of times that of ordinary full-color images (under the same spatial resolution conditions). This huge amount of data has brought a huge burden to the data storage and processing systems. Although the real-time processing method based on the hardware platform can improve the data processing speed to a certain extent, its research object is still the reconstructed three-dimensional data, and it cannot fundamentally solve the problem of large data volume that hinders the real-time application of hyperspectral. Summary of the invention:

[0006] The present invention aims to provide a hyperspectral rapid target detection method based on compressed spectral imaging to overcome the problem that the prior art cannot fundamentally solve the problem of large data volume that hinders the real-time application of hyperspectral.

[0007] In order to achieve the purpose of the present invention, a hyperspectral rapid target detection method based on compressed spectral imaging is provided by the present invention. The spectral coding template in the compressed spectral imaging technology is used to directly act the CEM operator on the radiation light of the spatial pixels entering the system, so that the system directly outputs the target detection measurement value of each spatial pixel.

[0008] The above method comprises the following steps:

[0009] Step 1: After the radiation light of a row of spatial pixels enters the hyperspectral target detection system through the slit, it is split into a two-dimensional spectrum by the dispersion element and projected onto the spatial light modulator.

[0010] Step 2: Control the modulation degree of the spatial light modulator according to the determined coding template, thereby modulating the spectrum after the current spatial row is split, wherein the coding template is obtained by the following steps:

[0011] Step 2.1: Calculate the autocorrelation matrix R according to formula (1)

[0012]

[0013] Among them, c i is the spectral vector of each ground feature sample in the reference spectral library, c i T It is c i is the transpose of , and N is the number of ground object categories in the reference spectral library.

[0014] Step 2.2: Based on the autocorrelation matrix R obtained in step 2.1 and the target spectrum d to be detected, determine the CEM operator w according to formula (2):

[0015]

[0016] Where w is a vector of size p×1, p represents the number of spectral channels, d is a spectral vector of the target to be detected of size p×1, and R is a sample autocorrelation matrix of size p×p.

[0017] Step 2.3: Expand the CEM operator w obtained in step 2.2 into a two-dimensional coding template along the spatial dimension according to the number of spatial pixels in a row;

[0018] Step 3: The modulated two-dimensional spectrum is reversely combined by another dispersion element and projected onto the detector to obtain the target detection measurement value of the current row space pixel.

[0019] Step 4: If the target space push scan is completed, the target detection is ended; otherwise, the next row of spatial pixels is pushed and scanned and the process goes to step 1.

[0020] Compared with the prior art, the advantages of the present invention are:

[0021] (1) The present invention utilizes the spectral encoding function of compressed spectral technology, combines the hyperspectral target detection algorithm with the compressed spectral imaging technology, and realizes a hyperspectral target detection system. The system can output the target detection results while modulating the spectrum without reconstructing the data cube. That is, the target detection research object is converted from three-dimensional spectral data to target space radiation light, and the target detection results are directly output while performing push scanning, thereby avoiding the problem of large data volume and providing a new solution for hyperspectral rapid target detection.

[0022] (2) For the target detection method provided by the present invention, when the application field and the target to be detected are clear, the autocorrelation matrix R and the target spectrum d can be determined in advance, and the CEM operator w and the coding template can also be determined in advance. Therefore, there is almost no time cost in the coding template generation process. On this basis, since the modulation speed of the spatial light modulator is very fast, even if the target space is pushed and scanned, the spatial target detection can be completed quickly. Description of the drawings:

[0023] Figure 1 It is a corresponding relationship diagram between target detection measurement values ​​of the present invention.

[0024] Figure 2 It is a flow chart of the present invention. Specific implementation method:

[0025] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0026] The system of the present invention is used to realize hyperspectral target detection, including a conventional spectral modulation optical path using compressed spectral imaging technology, a slit added at the incident light of the optical path, and a target detection coding template designed by the present invention, so that it can be transformed into a hyperspectral target detection system that projects target detection measurement values ​​onto a detector in a push-scan manner.

[0027] Constrained energy minimization (CEM) is a classic hyperspectral target detection method. The basic idea is to extract the signal (target) in a specific direction and attenuate the interference of signals (background) in other directions. The purpose of CEM is to design a FIR linear filter vector (also known as CEM operator) so that the output energy of the entire image is minimized after filtering, but the energy of the target is maximized after filtering, thereby highlighting the target and suppressing the background, so as to achieve the effect of separating the target from the entire image. The CEM operator w is,

[0028]

[0029] Where w is a vector of size p×1, p represents the number of spectral channels, d is a target spectral vector of size p×1, and R is a sample autocorrelation matrix of size p×p, which is generally counted over the entire spectral image range.

[0030] Compressed spectral imaging technology introduces a two-dimensional spectral coding template into dispersive spectral imaging, generates a sparse coding template based on the compressed sensing theory, performs spectral modulation on the two-dimensional aliased spectrum after splitting the radiation light entering the system, realizes spectral compression, and finally restores and reconstructs the spectrum obtained by compressed sampling according to the compressed sensing theory.

[0031] The present invention integrates the spectral coding template in the compressed spectral imaging technology and the CEM operator in the CEM target detection algorithm, and uses the spectral coding template in the compressed spectral imaging technology to directly act on the radiation light of the spatial pixels entering the system, so that the hyperspectral target detection system directly outputs the target detection measurement value of each spatial pixel. The present invention uses a slit to control each row of spatial pixels entering the hyperspectral target detection system, and realizes target detection in the entire two-dimensional space by a row-by-row push-scan method. Each time a row of spatial pixels entering the system radiates light, it is split and projected onto the spectral coding template, one dimension of which is the spatial dimension and one dimension is the spectral dimension. The autocorrelation matrix R is statistically calculated within the sample range of the reference spectral library,

[0032]

[0033] Among them, c i is the spectral vector of each ground feature sample in the reference spectral library, c i T It is c i is the transpose of , N is the number of ground object categories in the reference spectral library. The CEM operator w is calculated based on the statistical autocorrelation matrix R within the sample range of the reference spectral library, and w is used to generate the modulation coefficient matrix of the spectral coding template, that is, the vector direction of the CEM operator w is consistent with the spectral dimension direction of the spectral coding template, and w is expanded along the spatial dimension direction to cover the entire spectral coding template. The spectral coding template can use a spatial light modulator such as a digital micromirror array or a liquid crystal photodiode, and each unit of the spatial light modulator can be independently controlled by each modulation coefficient of the coding template. After the spectrum is modulated, the spectrum after modulating the spectral dimension direction is combined and projected onto the detector through the light combining module, and then the target detection measurement value corresponding to each spatial pixel is obtained on the detector. For each row of spatial pixels, the corresponding relationship between its spatial radiation light, the two-dimensional spectrum after splitting, the coding template, and the target detection measurement value obtained by the detector is as follows: Figure 1 shown.

[0034] Based on the above design principles, see Figure 2 The hyperspectral rapid target detection method based on compressed spectral imaging provided by the present invention comprises the following steps:

[0035] Step 1: After the radiation light of a row of spatial pixels enters the hyperspectral target detection system through the slit, it is split into a two-dimensional spectrum by the dispersion element and projected onto the spatial light modulator.

[0036] Step 2: Control the modulation degree of the spatial light modulator according to the determined coding template, thereby modulating the spectrum after the current spatial row is split, wherein the coding template is obtained by the following steps:

[0037] Step (1): Calculate the autocorrelation matrix R in the reference spectral library according to formula (1):

[0038]

[0039] Among them, c i is the spectral vector of each ground feature sample in the reference spectral library, c i T It is c i is the transpose of , and N is the number of ground object categories in the reference spectral library.

[0040] Step (2): Based on the autocorrelation matrix R obtained in step (1) and the target spectrum d to be detected, determine the CEM operator w according to formula (2):

[0041]

[0042] Where w is a vector of size p×1, p represents the number of spectral channels, d is a spectral vector of the target to be detected of size p×1, and R is a sample autocorrelation matrix of size p×p.

[0043] Step (3): Expand the CEM operator w obtained in step 2 into a two-dimensional coding template along the spatial dimension according to the number of spatial pixels in a row;

[0044] Step 3: The modulated two-dimensional spectrum is reversely combined by another dispersion element and projected onto the detector to obtain the target detection measurement value of the current row space pixel.

[0045] Step 4: If the target space push scan is completed, the target detection is ended. Otherwise, push scan to the next row of spatial pixels, repeat the above steps, and enter step 1.

[0046] Applying the CEM operator to all pixels in the spectral image can obtain the target detection measurement value of each pixel.

[0047] y i =w T x i (3)

[0048] where x i is the spectrum of the ith pixel (a vector of size p×1), y i is x i Corresponding target detection measurement value (1×1 size). Measurement value y i The larger the pixel x is, the i The greater the probability of being the target, and vice versa, that is, the measurement value presents the distribution of the target spectrum d in the entire spectral image.

Claims

1. A hyperspectral rapid target detection method based on compressed spectral imaging, characterized in that: It uses the spectral coding template in the compressed spectral imaging technology to directly apply the CEM operator to the radiation light of the spatial pixels entering the system, so that the system directly outputs the target detection measurement value of each spatial pixel; The specific steps include: Step 1: After the radiation light of a row of spatial pixels enters the hyperspectral target detection system through the slit, it is split into a two-dimensional spectrum by the dispersion element and projected onto the spatial light modulator; Step 2: Control the modulation degree of the spatial light modulator according to the determined coding template, thereby modulating the spectrum after the current spatial row is split, wherein the coding template is obtained by the following steps: Step 2.1: Calculate the autocorrelation matrix R according to formula (1) Among them, c i is the spectral vector of each ground feature sample in the reference spectral library, c i T It is c i is the transpose of , N is the number of ground object categories in the reference spectral library; Step 2.2: Based on the autocorrelation matrix R obtained in step 2.1 and the target spectrum d to be detected, determine the CEM operator w according to formula (2): Where w is a vector of size p×1, p represents the number of spectral channels, d is a spectral vector of the target to be detected of size p×1, and R is a sample autocorrelation matrix of size p×p; Step 2.3: Expand the CEM operator w obtained in step 2.2 into a two-dimensional coding template along the spatial dimension according to the number of spatial pixels in a row; Step 3: The modulated two-dimensional spectrum is reversely combined by another dispersion element and projected onto the detector to obtain the target detection measurement value of the current row space pixel; Step 4: If the target space push scan is completed, the target detection ends; otherwise, push scan to the next row of spatial pixels and go to step 1; Applying the CEM operator to all pixels in the spectral image can obtain the target detection measurement value of each pixel. y i =w T x i (3) where x i is the spectrum of the i-th pixel, a vector of size p×1, y i is x i Corresponding target detection measurement value, 1×1 size, measurement value y i The larger the pixel x is, the i The greater the probability of being the target, and vice versa, that is, the measurement value presents the distribution of the target spectrum d in the entire spectral image.

Images