A target detection method for snow-covered camouflaged targets

Abstract

The present invention discloses a target detection method for snow camouflage targets, which relates to the field of image detection, including: S1, collecting and preprocessing the spectra of typical snow and snow camouflage materials, and constructing an over-complete dictionary; S2, collecting and preprocessing airborne hyperspectral data; S3, calculating the difference band through the spectra of typical snow and snow camouflage materials, constructing the difference band selection constraint condition, and obtaining the background dictionary A after band screening based on the difference band b* , Joint dictionary A * The pixel y of the airborne hyperspectral image * ; S4, find the optimal sparse vector γ1 of the background dictionary and the optimal sparse vector β1 of the joint dictionary, and then find the background dictionary A b* The reconstruction error r0(y * ) and the joint dictionary A * The reconstruction error r1(y * ); S5, reconstruct the error r0(y * ) and reconstruction error r1(y * ) Through the SRUC detector, it is determined whether the target exists based on the output value. The present invention solves the problem that the existing target detection algorithm has a low detection rate for snow-camouflaged targets.

Description

Technical Field

[0001] The present invention relates to the field of image detection, and in particular to a target detection method for snow-camouflaged targets. Background Art

[0002] Hyperspectral target detection algorithms achieve target detection by mining the spectral difference information between the ground objects and the targets. Currently, the more widely used target detection algorithms include the Reed-XiaoLi algorithm (RX), the Constrained Energy Minimization algorithm (CEM), the Spectral Angle Matching algorithm (SAM), and the detection algorithm based on sparse representation.

[0003] Compared with typical target detection algorithms such as the RX algorithm and the CEM algorithm, in the target detection algorithm based on sparse representation, an over-complete dictionary composed of the spectral information of multiple targets and the background can be used to achieve target detection. However, at the same time, the purity of the over-complete dictionary composed of the target and the background also has a great impact on the detection result. At the present stage, the target dictionary is usually selected manually by visual inspection, and the background dictionary is selected by the method of visually setting concentric double windows.

[0004] The existing target detection algorithms can be competent for the inspection tasks of typical targets based on sparse representation. However, for the camouflaged targets hidden by snow-camouflaged materials with extremely low spectral differences from the snow background, there will be problems of poor detection effects due to insufficient mining of spectral information.

[0005] In the process of target detection by the target detection algorithm based on sparse representation, due to the very small visual difference between the snow-camouflaged target and the snow background, impurities are easily mixed into the target over-complete dictionary composed of the spectra of the visually selected snow-camouflaged targets and the over-complete background dictionary selected by setting concentric double windows, which has a great impact on the detection result. Summary of the Invention

[0006] The present invention aims to provide a target detection method for snow-camouflaged targets to solve the problem of poor detection effects due to insufficient mining of spectral information for snow-camouflaged targets with extremely low spectral differences from the snow background.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A target detection method for snow-camouflaged targets, the target detection method for snow-camouflaged targets includes the following steps:

[0009] S1. Acquisition and preprocessing of the spectra of typical snow and snow-camouflaged materials, construction of a background dictionary A based on the spectra of the preprocessed typical snow b , and construction of a combined dictionary A based on the spectra of the preprocessed typical snow and snow-camouflaged materials;

[0010] S2. Acquisition and preprocessing of airborne hyperspectral data;

[0011] S3. Calculate the differential bands between the snow camouflage material and the typical snow, construct the constraint conditions for selecting differential bands, and based on the differential bands, for the background dictionary A b 、Joint the dictionary A and the airborne hyperspectral image for band screening to obtain the background dictionary A after band screening b* 、Joint the dictionary A * with the pixel y of the airborne hyperspectral image * ;

[0012] S4. Based on the background dictionary A after band screening b* 、Joint the dictionary A * with the pixel y of the airborne hyperspectral image * , use the orthogonal matching pursuit algorithm to find the optimal sparse vector γ of the background dictionary 1 and the optimal sparse vector β of the joint dictionary 1 , then calculate the reconstruction error r b* of the background dictionary A for the pixels of the image after band screening 0 (y * ) and the reconstruction error r * of the joint dictionary A for the pixels of the image after band screening 1 (y * );

[0013] S5. Pass the reconstruction error r 0 (y * ) and the reconstruction error r 1 (y * ) through the SRUC detector to obtain the output value of the SRUC model, and judge whether the target exists based on the output result.

[0014] Further, in step S1, the spectra of the typical snow and the snow camouflage material are preprocessed by filtering denoising and spectral resampling.

[0015] Further, in step S1, after preprocessing the spectra of the typical snow and the snow camouflage material, spectral data with a spectral resolution of 4 nm and a wavelength range of 400 - 1000 nm are obtained. There are 150 spectra of the typical snow and 70 spectra of the snow camouflage material.

[0016] Further, in step S2, the airborne hyperspectral data is preprocessed by registration, radiometric correction, geometric correction, filtering denoising, and spectral resampling.

[0017] Further, in step S2, the wavelength range of the preprocessed airborne hyperspectral image is 400 - 1000 nm, the band interval is 4 nm, and there are 150 bands.

[0018] Further, in step S3, the differential band calculation formula is:

[0019]

[0020] where α ( u i , v i) is the included angle between the vector of the spectral curve u of the camouflage net and the spectral curve v of the snow in the i-th and (i + 1)-th bands;

[0021] The constraint condition for selecting the differential band is:

[0022] (i, i + 1) = f(α > α max )

[0023] where α max represents the band with a high difference between the snow camouflage net and the snow spectrum, and (i, i + 1) represents the differential band selected under the constraint condition for selecting the differential band.

[0024] Further, the α value is taken as the top 40% of the cumulative percentage.

[0025] Further, in step S4, the optimal sparse vector γ 1 of the background dictionary is:

[0026]

[0027] In the formula, γ represents the sparse vector of the background dictionary, and L represents the sparsity;

[0028] The optimal sparse vector β 1 of the combined dictionary is:

[0029]

[0030] In the formula, β represents the sparse vector of the combined dictionary, and L represents the sparsity.

[0031] Further, in step S4, the reconstruction error r b* of the background dictionary A 0 for the pixel of the image after band screening (y * ) is:

[0032] r 0 (y * ) = ||y * - A b* γ 1 ||

[0033] The reconstruction error r * of the combined dictionary A 1 for the pixel of the image after band screening (y * ) is:

[0034] r 1 (y * )=||y * -A * β 1 ||。

[0035] Furthermore, in step S5, the formula of the SRUC detector is:

[0036] D SRUC =r 0 (y * )-r 1 (y * )。

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention proposes a constraint condition for selecting the difference band according to the difference band between the camouflage target and the snow background, and improves the target detection method of sparse representation; and constructs a dictionary based on the spectral composition of typical snow and snow camouflage materials measured in the laboratory to ensure the purity of the dictionary, and solves the problem of low detection rate of the existing target detection algorithm for snow camouflage targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flowchart of a target detection method for snow camouflage targets. DETAILED DESCRIPTION OF THE INVENTION

[0040] The present invention will be further described in detail below with reference to the drawings and embodiments:

[0041] As Figure 1 shown, a target detection method for snow camouflage targets includes the following steps:

[0042] S1. Acquisition, preprocessing of the spectra of typical snow and snow camouflage materials, and construction of an over-complete dictionary;

[0043] Spectral measurements are taken on typical snow and snow camouflage materials to obtain two spectral trends. A total of 150 typical snow sample spectra and 70 snow camouflage material sample point spectra are obtained. After filtering and denoising, spectral resampling processing, spectral data with a spectral resolution of 4 nm and a wavelength range of 400 - 1000 nm is obtained, and a background dictionary A b is constructed based on the preprocessed typical snow spectra, and a joint dictionary A is constructed based on the preprocessed typical snow and snow camouflage material spectra.

[0044] S2. Acquisition and preprocessing of airborne hyperspectral data;

[0045] Airborne hyperspectral data of snow camouflage targets deployed in a snow environment are collected. After registration, radiometric correction, geometric correction, filtering and denoising, and spectral resampling, hyperspectral images with a spectral wavelength range of 400 - 1000 nm, a band interval of 4 nm, and 150 bands are obtained.

[0046] S3. Calculate the differential bands by using the typical snow spectrum and the snow camouflage material spectrum to construct the constraint conditions for selecting differential bands. The formula for the differential bands is shown in Equation 1.

[0047]

[0048] In the formula, α ( u i , v i) is the included angle between the vector of the snow camouflage material spectral curve u and the typical snow spectral curve v in the i-th and (i + 1)-th bands.

[0049] Then, establish the constraint condition for selecting differential bands (i, i + 1) = f(α > α max ), where α max represents the bands with a large difference between the snow camouflage material and the typical snow spectrum. Take the α values with a cumulative percentage of the top 40%, and (i, i + 1) represents the differential bands selected under the constraint condition for selecting differential bands. Refer to the differential bands between the snow camouflage material and the typical snow spectrum to screen the background dictionary A b , the combined dictionary A and the airborne hyperspectral image to obtain the background dictionary A b* , the combined dictionary A * and the airborne hyperspectral image after band screening.

[0050] S4. Based on the screened background dictionary A b* , the combined dictionary A * and the pixel y * of the airborne hyperspectral image, solve to obtain the optimal coefficient γ 1 of the sparse vector of the background dictionary and the optimal coefficient β 1 of the sparse vector of the combined dictionary, as shown in Equation 2 and Equation 3.

[0051]

[0052] In Equation 2 and Equation 3, γ represents the sparse vector of the background dictionary, β represents the sparse vector of the combined dictionary, and L represents the sparsity.

[0053] After obtaining the optimal sparse vector γ 1 of the background dictionary and the optimal sparse vector β 1 of the combined dictionary through the orthogonal matching pursuit algorithm, use Equation 4 and Equation 5 to calculate the reconstruction error r b* of the background dictionary A for the pixels of the image after band screening respectively.0 (y * ) and the combined dictionary A * The reconstruction error r of the image pixels after band screening 1 (y * ).

[0054] r 0 (y * ) = ||y * - A b* γ 1 || (4)

[0055] r 1 (y * ) = ||y * - A * β 1 || (5)

[0056] S5. Pass the reconstruction error r 0 (y * ) and the reconstruction error r 1 (y * ) through the SRUC detector, whose formula is shown in Equation 6, to obtain the output value of the SRUC model, and judge whether the target exists based on the output result.

[0057] D SRUC = r 0 (y * ) - r 1 (y * ) (6)

[0058] According to the specific examples provided by the present invention, the present invention discloses the following technical effects: The AUC value for snow camouflage environment detection reaches 0.92, and the detection effect on snow camouflage materials is better than that of RX, CEM, SAM, and traditional sparse representation algorithms.

[0059] The above are only the embodiments of the present invention. Specific technical solutions and / or common knowledge such as characteristics known in the solutions are not described in detail here. It should be noted that for those skilled in the art, without departing from the technical solution of the present invention, several deformations and improvements can still be made, and these should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners described in the specification can be used to explain the content of the claims.

Claims

1. A target detection method for a camouflaged target on snow, characterized in that: The target detection method for a snow-camouflaged target comprises the following steps: S1. Collection and preprocessing of spectra of typical snow and snow camouflage materials, and construction of background dictionary A based on the preprocessed spectra of typical snow b , construct a joint dictionary A based on the spectra of typical snow and snow camouflage materials after preprocessing; S2, collection and preprocessing of airborne hyperspectral data; S3, calculate the difference bands through the spectra of typical snow and snow camouflage materials, construct the difference band selection constraint conditions, and select the background dictionary A based on the difference bands b , perform band screening on the combined dictionary A and the airborne hyperspectral image to obtain the background dictionary A after band screening b* , Joint dictionary A * The pixel y of the airborne hyperspectral image * ; S4, background dictionary A based on band-based band screening b* , Joint dictionary A * The pixel y of the airborne hyperspectral image * , the optimal sparse vector γ1 of the background dictionary and the optimal sparse vector β1 of the joint dictionary are obtained by the orthogonal matching pursuit algorithm, and then the background dictionary A is obtained b* The reconstruction error r0(y * ) and the joint dictionary A * The reconstruction error r1(y * ); S5, reconstruct the error r0(y * ) and reconstruction error r1(y * ) through the SRUC detector, the formula of the SRUC detector is: D SRUC =r0(y * )-r1(y * ) Get the output value of the SRUC model and determine whether the target exists based on the output result.

2. The target detection method for a camouflaged target on snow according to claim 1, characterized in that: In step S1, the spectra of typical snow and snow camouflage materials are pre-processed by filtering, denoising and spectrum resampling.

3. The target detection method for a camouflaged target on snow according to claim 2, characterized in that: In step S1, after preprocessing, the spectra of typical snow and snow camouflage materials are used to obtain spectral data with a spectral resolution of 4nm and a wavelength range of 400-1000nm. There are 150 spectra of typical snow and 70 spectra of snow camouflage materials.

4. The target detection method for a snow-camouflaged target according to claim 1, characterized in that: In step S2, the airborne hyperspectral image is preprocessed by registration, radiation correction, geometric correction, filtering and denoising, and spectral resampling.

5. The target detection method for a camouflaged target on snow according to claim 4, characterized in that: In step S2, the preprocessed airborne hyperspectral image has a wavelength range of 400-1000 nm, a band interval of 4 nm, and 150 bands.

6. The target detection method for a snow-camouflaged target according to claim 1, characterized in that: In step S3, the difference band calculation formula is: In the formula, α(u i ,v i ) is the angle between the spectral curve u of the camouflage material and the spectral curve v of the typical snow field in the i and i+1 bands; The constraints for selecting difference bands are: (i,i+1)=f(α>α max ) where α max It represents the band with high spectrum difference between snow camouflage material and typical snow, and (i,i+1) represents the difference band selected under the constraint condition of difference band selection.

7. The target detection method for a camouflaged target on snow according to claim 6, characterized in that: The α value is taken as the first 40% of the cumulative percentage.

8. The target detection method for a camouflaged target on snow according to claim 1, characterized in that: In step S4, the optimal sparse vector γ1 of the background dictionary is: In the formula, γ represents the sparse vector of the background dictionary, and L represents the sparsity; The optimal sparse vector β1 of the joint dictionary is: Where β represents the sparse vector of the joint dictionary and L represents the sparsity.

9. The target detection method for a camouflaged target on snow according to claim 1, characterized in that: In step S4, the background dictionary A b* The reconstruction error r0(y * )for: r0(y * )=||y * -A b* γ1|| Joint Dictionary A * The reconstruction error r1(y * )for: r1(y * )=||y * -A * β1||。

Images