A real-time spectral target detection method using a snapshot spectral camera

Abstract

A real-time spectral target detection method using a snapshot spectral camera, relating to the field of target detection, alleviates the problem of long processing times in existing spectral target detection algorithms, which cannot meet the real-time spectral target detection requirements of snapshot spectral cameras. The method involves acquiring a spectral image of the region containing the target to be detected; performing dimensionality reduction on the spectral image to obtain a three-band pseudo-color image; the spectral image includes spectral images from multiple channels; performing salient region detection on the three-band pseudo-color image to obtain salient region detection results; performing edge contour detection on the three-band pseudo-color image to obtain edge contour detection results; and weighted fusing the salient region detection results and the edge contour detection results to obtain the spectral target detection result. This method is applicable to real-time spectral target detection of spectral images acquired by a snapshot spectral camera.

Description

Technical Field

[0001] This invention relates to the field of target detection and also to the field of spectral imaging technology. Background Technology

[0002] Spectral imaging technology, as an important remote sensing observation method, not only collects spatial feature information of objects in a scene, but also captures spectral feature information of targets at different wavelengths, and has been widely used in many fields such as environmental protection, agricultural monitoring, and medical imaging. Snapshot-type spectroscopic cameras, employing multispectral filter arrays, are an emerging spectroscopic imaging technology that combines the advantages of multispectral imaging and snapshot imaging. They overcome the limitations of sequential acquisition in traditional spectroscopic imaging, enabling the simultaneous capture of images across multiple spectral bands in a single exposure time, making real-time spectroscopic imaging a reality.

[0003] Spectral target detection is an important research direction in the field of spectral imaging. By utilizing spectral imaging technology to analyze the spectral characteristics of a target object at different wavelengths, it is possible to distinguish the target from the background, thereby achieving high-precision target detection. Snapshot spectral cameras have made significant progress in spectral image acquisition efficiency, but existing spectral target detection algorithms are complex and have long data processing times, making it difficult to meet the needs of real-time processing.

[0004] In summary, existing spectral target detection algorithms suffer from long processing times and cannot meet the requirements for real-time spectral target detection by snapshot spectral cameras. Summary of the Invention

[0005] This invention alleviates the problem that existing spectral target detection algorithms have long processing times and cannot meet the real-time spectral target detection requirements of snapshot spectral cameras. This invention provides the following solution:

[0006] Option 1: A spectral target detection method, comprising the following steps:

[0007] Step S1: Acquire the spectral image of the region where the target to be detected is located, and perform dimensionality reduction on the spectral image to obtain a three-band pseudo-color image; the spectral image includes spectral images of multiple channels;

[0008] Step S2: Perform salient region detection on the three-band pseudo-color image to obtain the salient region detection result;

[0009] Step S3: Perform edge contour detection on the three-band pseudo-color image to obtain the edge contour detection result;

[0010] Step S4: The salient region detection results and edge contour detection results are weighted and fused to obtain the spectral target detection results.

[0011] Furthermore, in one embodiment of the present invention, step S1 includes:

[0012] Step S11: Select the spectral images of three channels from the spectral image;

[0013] Step S12: Based on the spectral images of the three channels, obtain a three-band pseudo-color image.

[0014] Furthermore, in one embodiment of the present invention, step S11 includes:

[0015] Step S111: For the spectral image of each channel in the spectral image, obtain its corresponding spectral average response image;

[0016] Step S112: Based on the average response images of all spectra, obtain the spectral difference score of the spectral image for each channel;

[0017] Step S113: Sort the spectral images of each channel according to the spectral difference score, and select the spectral images of the three channels with the highest spectral difference scores.

[0018] Furthermore, in one embodiment of the present invention, the spectral difference fraction mentioned in step S112... pass

[0019]

[0020] Obtained, wherein the spectral image contains * 1 pixel; For the first The coordinates in the spectral channel image of each channel are The grayscale value of the pixel; The coordinates in the spectral average response image are The grayscale value of the pixel.

[0021] Furthermore, in one embodiment of the present invention, the spectral difference fraction mentioned in step S112 is... pass

[0022]

[0023] get.

[0024] Furthermore, in one embodiment of the present invention, step S2 is:

[0025] Step A21: Perform a Fourier transform on the three-band pseudo-color image to obtain a complex matrix, wherein the complex matrix includes an amplitude spectrum and a phase spectrum;

[0026] Step A22: Perform a logarithmic transformation on the amplitude spectrum to obtain a logarithmic amplitude spectrum; filter the logarithmic amplitude spectrum to obtain a smoothed amplitude spectrum;

[0027] The filtering method used is either mean filtering or Gaussian filtering.

[0028] Step A23: Subtract the smoothed amplitude spectrum obtained in step A22 from the amplitude spectrum obtained in step A21 to obtain the spectral residual;

[0029] Step A24: Combine the phase spectrum and the spectral residual and perform an inverse Fourier transform to obtain the salient region detection result.

[0030] Furthermore, in one embodiment of the present invention, step S2 is:

[0031] Step B21: Convert the three-band pseudo-color image to the XYZ color space to obtain the XYZ pseudo-color image;

[0032] Step B22: Convert the XYZ pseudo-color image to the LAB color space to obtain the LAB color image;

[0033] Step B23: Based on the LAB color image, obtain the LAB color vector of each pixel and the average LAB color vector of all pixels;

[0034] Step B24: The salient region detection result is obtained based on the Euclidean distance between the LAB color vector of each pixel and the average LAB color vector of the image.

[0035] Furthermore, in one embodiment of the present invention, step S3 is:

[0036] Step A31: Perform the Canny edge detection algorithm on the three-band pseudo-color image to obtain the edge results of the three channels in the three-band pseudo-color image; accumulate and perform binarized threshold segmentation on the image edge results of the three channels to obtain preliminary edge contour results;

[0037] Step A32: Perform image opening operation on the obtained preliminary edge contour results to obtain the edge contour detection results.

[0038] Furthermore, in one embodiment of the present invention, step S3 is:

[0039] Step B31: The gradient magnitudes of the three channels in the three-band pseudo-color image are obtained by using the Sobel algorithm.

[0040] Step B32: Accumulate the gradient magnitudes of the three channels and perform binarization thresholding to obtain preliminary edge contour results;

[0041] The gradient magnitude includes the horizontal gradient magnitude and the vertical gradient magnitude;

[0042] Step B33: Perform image opening operation on the obtained preliminary edge contour results to obtain the edge contour detection results.

[0043] Furthermore, in one embodiment of the present invention, the weights of the weighted fusion in step S4 are all set to 0.5.

[0044] The spectral target detection method described in this invention is based on fusion detection, which effectively alleviates the problem that existing spectral target detection algorithms have long processing times and cannot meet the real-time spectral target detection requirements of snapshot spectral cameras. Specific beneficial effects include:

[0045] 1. In the spectral target detection method of the present invention, the dimensionality reduction of the spectral image is achieved based on spectral differences. Existing snapshot spectral camera spectral image data consists of multi-band images. These high-dimensional image data not only increase computational complexity but may also induce the Hughes phenomenon, posing a significant challenge to real-time target detection. Therefore, to improve the efficiency and accuracy of spectral target detection, the present invention selects the most representative channel spectral image based on spectral differences to reduce the dimensionality of the spectral image into a three-band pseudo-color image. This significantly reduces the number of channels while preserving effective data for target identification, thereby ensuring the accuracy of target identification and significantly improving detection speed. Compared with existing similar technologies, the method described in the present invention improves the detection speed by 37-1576 times, meeting the requirements of real-time processing and suitable for the speed requirements of spectral image processing in snapshot spectral cameras.

[0046] 2. In the spectral target detection method described in this invention, the spectral image is dimensionality reduced, which decreases the amount of data processing and improves the detection speed. To ensure detection accuracy while reducing the amount of data, this invention integrates the results of salient region detection and edge contour detection, ensuring more comprehensive target capture with limited data, and obtaining clearer target regions. This effectively reduces the probability of missed and false detections while ensuring detection accuracy.

[0047] The method described in this invention is applicable to real-time spectral target detection of spectral images obtained by a snapshot spectral camera. Attached Figure Description

[0048] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0049] Figure 1This is a flowchart of the spectral target detection method described in Implementation Method 1;

[0050] Figure 2 This is a schematic diagram of the imaging process of the snapshot spectral camera described in Embodiment 1;

[0051] Figure 3 This is a flowchart of step S1, optimized in embodiment four, step S2, optimized in embodiment six, and step S3, optimized in embodiment eight, as described in embodiment eleven.

[0052] Figure 4 This is a schematic diagram of the filter array channel described in Embodiment Eleven;

[0053] Figure 5 It is the spectral image of the region where the large target to be detected is located, as described in Implementation Method Eleven;

[0054] Figure 6 These are comparison charts showing the effects of large target detection as described in Embodiment Eleven, where (a) is the detection effect using the existing LRX algorithm, (b) is the detection effect using the existing CRD algorithm, (c) is the detection effect using the existing LRASR algorithm, (d) is the detection effect using the existing LSMAD algorithm, (e) is the detection effect using the existing PTA algorithm, (f) is the detection effect using the existing SSIIFD algorithm, (g) is the detection effect of the target detection method described in Embodiment Eleven, and (h) is the detection effect of the target detection method described in Embodiment Twelve.

[0055] Figure 7 It is the spectral image of the region where the small target to be detected is located, as described in Implementation Method Eleven;

[0056] Figure 8 These are comparison diagrams of the small target detection results described in Embodiment Eleven, wherein (a) is the detection result using the existing LRX algorithm, (b) is the detection result using the existing CRD algorithm, (c) is the detection result using the existing LRASR algorithm, (d) is the detection result using the existing LSMAD algorithm, (e) is the detection result using the existing PTA algorithm, (f) is the detection result using the existing SSIIFD algorithm, (g) is the detection result of the target detection method described in Embodiment Eleven, and (h) is the detection result of the target detection method described in Embodiment Twelve.

[0057] Figure 9 yes Figure 6 , Figure 8 A comparison chart of AUC values ​​for each of the shown renderings;

[0058] Figure 10 yes Figure 6 , Figure 8A comparison chart of image processing times for each of the shown renderings;

[0059] Figure 11 This is a flowchart of the combination of steps S1 (optimized from implementation method 5), S2 (optimized from implementation method 7), and S3 (optimized from implementation method 9) as described in implementation method 12. Detailed Implementation

[0060] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0061] Implementation Method 1: The spectral target detection method described in this implementation method includes the following steps, such as... Figure 1 As shown:

[0062] Step S1: Acquire the spectral image of the region where the target to be detected is located, and perform dimensionality reduction on the spectral image to obtain a three-band pseudo-color image; the spectral image includes spectral images of multiple channels;

[0063] Step S2: Perform salient region detection on the three-band pseudo-color image to obtain the salient region detection result;

[0064] Step S3: Perform edge contour detection on the three-band pseudo-color image to obtain the edge contour detection result;

[0065] Step S4: The salient region detection results and edge contour detection results are weighted and fused to obtain the spectral target detection results.

[0066] The preferred hardware platform for this embodiment is a desktop computer with a 64-bit Windows 10 operating system, a 12th Gen Intel(R) Core(TM) i5-12400F @ 2.50 GHz processor, and 16GB of RAM; the implementation is carried out in the MATLAB R2022b software environment.

[0067] This embodiment preferably performs real-time spectral target detection on the spectral images obtained by the snapshot spectral camera, and the imaging process of the snapshot spectral camera is as follows: Figure 2 As shown, light emitted from the light source passes through the target scene, and the reflected light is focused onto the camera image sensor through the lens. The image sensor array is covered with a 9-channel multispectral filter array, capturing a 9-channel spectral image of the target scene.

[0068] The spectral target detection method described in this embodiment first performs dimensionality reduction processing on the spectral image of the target region, reducing the amount of data processing and thus improving the image processing speed.

[0069] Furthermore, this embodiment also achieves real-time spectral target detection based on a saliency and edge contour fusion technique. This technique fuses the detection results of saliency region detection and edge contour detection, combining them with the preceding spectral dimensionality reduction to form an innovative and efficient real-time spectral target detection workflow. This significantly improves processing speed and detection accuracy, meeting the real-time application requirements of snapshot-type spectroscopic cameras. It effectively alleviates the problem of long processing times in existing spectral target detection algorithms, which cannot meet the real-time spectral target detection needs of snapshot-type spectroscopic cameras.

[0070] Implementation Method Two: This implementation method further defines the spectral target detection method described in Implementation Method One. In this implementation method, step S1 includes:

[0071] Step S11: Select the spectral images of three channels from the spectral image;

[0072] Step S12: Based on the spectral images of the three channels, obtain a three-band pseudo-color image.

[0073] This embodiment further defines step S1 and provides an example of the solution in step S1. Since the spectral image data of a snapshot spectroscopic camera consists of multi-band images, this high-dimensional image data not only increases computational complexity but may also trigger the Hughes phenomenon, posing a significant challenge to real-time target detection. To improve the efficiency and accuracy of spectral target detection, based on the differences in spectral images, three-channel spectral images are selected to reduce the dimensionality of the spectral image to be detected into a three-band pseudo-color image. This significantly reduces the number of bands while maintaining accuracy. This method effectively reduces the dimensionality of hyperspectral images by combining three-channel spectral images into a new three-band pseudo-color image, balancing information preservation and computational efficiency, laying the foundation for subsequent real-time processing.

[0074] Implementation Method 3: This implementation method further defines the spectral target detection method described in Implementation Method 2. In this implementation method, step S11 includes:

[0075] Step S111: For the spectral image of each channel in the spectral image, obtain its corresponding spectral average response image;

[0076] Step S112: Based on the average response images of all spectra, obtain the spectral difference score of the spectral image for each channel;

[0077] Step S113: Sort the spectral images of each channel according to the spectral difference score, and select the spectral images of the three channels with the highest spectral difference scores.

[0078] In this embodiment, step S111 is achieved through...

[0079]

[0080] Obtain the spectral average response image , In the spectral average response image The grayscale value of the pixel at the location, where, For the number of channels, For the first In the spectral channel image of the channel The grayscale value of the pixel at the location.

[0081] This method quantifies the spectral differences between different spectral channels by calculating the mean absolute error between each spectral channel image and the average spectral response image. Mean absolute error, as a commonly used error assessment metric, can effectively reflect the differences and information content of spectral channels.

[0082] This embodiment further defines step S11 and provides an example of the technical means of step S11. This method quantifies the differences between spectral channels and effectively filters out the most informative spectral channels based on the differences, avoiding the information loss problem that may occur in traditional dimensionality reduction methods. It also ensures the accuracy and precision of target detection even after dimensionality reduction of the spectral image.

[0083] The three spectral channels with the highest difference scores are selected because a higher difference score indicates greater spectral difference, and vice versa. Higher difference scores indicate significant differences between the spectral channels and the mean image, suggesting more useful information. Therefore, this method preserves key information in the spectral images, improving the efficiency of subsequent analysis and processing.

[0084] Implementation Method Four: This implementation method further defines the spectral target detection method described in Implementation Method Three. In this implementation method, the spectral difference fraction mentioned in step S112... pass

[0085]

[0086] Obtained, wherein the spectral image contains * 1 pixel; For the first The coordinates in the spectral channel image of each channel are The grayscale value of the pixel; The coordinates in the spectral average response image are The grayscale value of the pixel.

[0087] This embodiment further defines S112 and illustrates the technical means of S112. This method uses the mean absolute error (MAE) method to measure the difference between the average spectral response image and the image of each individual spectral channel, obtaining a difference score for each channel. The MAE, as a commonly used error evaluation index, can effectively reflect the differences and information content of spectral channels.

[0088] Implementation Method Five: This implementation method further defines the spectral target detection method described in Implementation Method Three. In this implementation method, the spectral difference score mentioned in step S112... pass

[0089]

[0090] get.

[0091] This embodiment further defines S1 and provides an example of the technical means of S1. The method uses mean squared error (MSE) to measure the difference between the average spectral response image and each independent spectral channel image, and obtains the difference score of each channel, which can effectively reflect the difference and information content of the spectral channels.

[0092] Implementation Method Six: This implementation method further defines the spectral target detection method described in Implementation Method One. In this implementation method, step S2 is as follows:

[0093] Step A21: Perform a Fourier transform on the three-band pseudo-color image to obtain a complex matrix, wherein the complex matrix includes an amplitude spectrum and a phase spectrum;

[0094] Step A22: Perform a logarithmic transformation on the amplitude spectrum to obtain a logarithmic amplitude spectrum; filter the logarithmic amplitude spectrum to obtain a smoothed amplitude spectrum;

[0095] The filtering method used is either mean filtering or Gaussian filtering.

[0096] Step A23: Subtract the smoothed amplitude spectrum obtained in step A22 from the amplitude spectrum obtained in step A21 to obtain the spectral residual;

[0097] Step A24: Combine the phase spectrum and the spectral residual and perform an inverse Fourier transform to obtain the salient region detection result.

[0098] This embodiment further defines step S2 and provides an example of the technical means used in step S2. This method utilizes a residual spectrum algorithm to detect salient regions in the obtained three-band pseudo-color image. By performing a logarithmic transformation on the amplitude spectrum, it enhances low-grayscale details and compresses the high-grayscale range. By filtering the obtained logarithmic amplitude spectrum, it removes background information while preserving edge and texture features. The frequency domain logarithmic amplitude spectra of different images exhibit similar distribution trends and satisfy local linearity conditions. This similarity reflects redundant information in the three-band pseudo-color image. Removing this redundant information yields salient information that attracts the attention of the human visual system, effectively identifying salient regions in the image and thus achieving target-background segmentation.

[0099] This implementation method works better when used in conjunction with the methods in implementation methods two to five.

[0100] Implementation Method Seven: This implementation method further defines the spectral target detection method described in Implementation Method One. In this implementation method, step S2 is as follows:

[0101] Step B21: Convert the three-band pseudo-color image to the XYZ color space to obtain the XYZ pseudo-color image;

[0102] Step B22: Convert the XYZ pseudo-color image to the LAB color space to obtain the LAB color image;

[0103] Step B23: Based on the LAB color image, obtain the LAB color vector of each pixel and the average LAB color vector of all pixels;

[0104] Step B24: The salient region detection result is obtained based on the Euclidean distance between the LAB color vector of each pixel and the average LAB color vector of the image.

[0105] In this embodiment, obtaining the salient region detection result based on the Euclidean distance involves identifying pixels with larger Euclidean distances as salient regions, thus obtaining the salient region detection result. Since a larger Euclidean distance indicates a greater difference, the pixel appears brighter and more prominent in the image. Therefore, those skilled in the art can make this judgment based on the characteristics of the three-band pseudo-color image to be detected.

[0106] This embodiment further defines step S2 and illustrates the technical means of step S2. The method is a saliency detection algorithm based on the Euclidean distance between pixel vectors and the average pixel vector in the LAB color space. In this method, the LAB color space, due to its better color consistency and uniform perceptual characteristics, can provide more reliable color features for subsequent saliency detection. Replacing the traditional RGB space of a three-band pseudo-color image with the LAB color space improves the accuracy of color feature expression and enhances the simulation effect of visual saliency. Since the RGB color space cannot be directly converted to the LAB color space, it requires the use of the XYZ color space. The RGB color space must first be converted to the XYZ color space, and then the XYZ color space is converted to the LAB color space. Therefore, this method ensures the scientific nature and accuracy of color conversion through a two-step conversion (RGB→XYZ→LAB).

[0107] Then, by using the Euclidean distance of pixels in LAB space as a saliency measure, the sensitivity of human vision to color differences is simulated, which can effectively identify salient regions in the image, achieve target and background segmentation, improve the accuracy and robustness of saliency detection, and break through the limitations of traditional reliance on single low-level features such as color, brightness and texture.

[0108] This implementation method works better when used in conjunction with the methods in implementation methods two to five.

[0109] Implementation Method Eight: This implementation method further defines the spectral target detection method described in Implementation Method One or Implementation Method Six. In this implementation method, step S3 is as follows:

[0110] Step A31: Perform the Canny edge detection algorithm on the three-band pseudo-color image to obtain the edge results of the three channels in the three-band pseudo-color image; accumulate and perform binarized threshold segmentation on the image edge results of the three channels to obtain preliminary edge contour results;

[0111] Step A32: Perform image opening operation on the obtained preliminary edge contour results to obtain the edge contour detection results.

[0112] In this embodiment, the image opening operation specifically involves performing a connected component operation on the preliminary edge contour result to remove bright areas that are less than a certain threshold.

[0113] This embodiment further defines step S3 and illustrates the technical means of step S3 with examples. This method utilizes the Canny edge detection algorithm and image opening operation for edge contour detection. Considering the computational load and processing efficiency of the edge detection algorithm, as well as the requirements of real-time processing, the Canny edge detection algorithm is used for image processing. This can effectively identify the edge contours of objects and also has a noise smoothing effect. Since the detection results of the Canny edge detection algorithm may contain false detection areas, this method also combines the Canny edge detection algorithm with image opening operation for edge contour detection, effectively balancing detection accuracy and real-time performance. This is a targeted optimization for the real-time requirements of snapshot spectral cameras, ensuring the integrity of edge information and minimizing noise, effectively balancing detection accuracy and real-time performance.

[0114] This implementation method works better when used in conjunction with the methods of implementation methods two to five and implementation method six.

[0115] Implementation Method Nine: This implementation method further defines the spectral target detection method described in Implementation Method One or Implementation Method Seven. In this implementation method, step S3 is as follows:

[0116] Step B31: The gradient magnitudes of the three channels in the three-band pseudo-color image are obtained by using the Sobel algorithm.

[0117] Step B32: Accumulate the gradient magnitudes of the three channels and perform binarization thresholding to obtain preliminary edge contour results;

[0118] The gradient magnitude includes the horizontal gradient magnitude and the vertical gradient magnitude;

[0119] Step B33: Perform image opening operation on the obtained preliminary edge contour results to obtain the edge contour detection results.

[0120] In this embodiment, the image opening operation specifically involves performing a connected component operation on the preliminary edge contour result to remove bright areas that are less than a certain threshold.

[0121] This embodiment further defines step S3 and provides an example of the technical means employed in step S3. This method utilizes gradient calculation of the Sobel operator and image opening operations for edge contour detection. Specifically, the Sobel operator is used to calculate the horizontal and vertical gradients of the image, effectively identifying the edge contours of objects while smoothing noise. Based on this, the false detection elimination through image opening operations effectively balances detection accuracy and real-time performance. This step is specifically optimized for the real-time requirements of snapshot-type spectral cameras, ensuring the integrity of edge information and minimizing noise.

[0122] This implementation method works better when used in conjunction with the methods of implementation methods two to five and implementation method seven.

[0123] Implementation Method 10: This implementation method further defines the spectral target detection method described in Implementation Method 1. In this implementation method, the weights of the weighted fusion in step S4 are all set to 0.5.

[0124] In this embodiment, by

[0125]

[0126] Obtain spectral target detection results ,in, The weighted fusion coefficient is the result of salient region detection. This represents the weighted fusion coefficient of the edge contour detection results.

[0127] This embodiment further defines step S4 and provides an example of the scheme in step S4. This method weightedly fuses the obtained salient region detection results and edge contour detection results, combining the advantages of both detection methods to improve the accuracy and robustness of target detection. A simple and efficient weighted fusion strategy is designed, fully utilizing the complementary information of salient regions and edge contours. This ensures computational efficiency while significantly improving detection performance, capturing targets more comprehensively, making the target area clearer, reducing the probability of missed and false detections, and meeting the needs of real-time spectral target detection.

[0128] Implementation Method Eleven: This implementation method compares the spectral target detection method of the present invention with existing similar technologies by conducting target detection experiments on the same large target and small target, and then compares the detection results to prove that the spectral target detection method of the present invention has a faster detection speed and better detection accuracy.

[0129] The spectral target detection method used in this embodiment is based on the spectral target detection method described in Embodiment 1, combined with the optimized steps S1 of Embodiment 4, S2 of Embodiment 6, and S3 of Embodiment 8. The flowchart is as follows. Figure 3 As shown.

[0130] Existing similar technologies have selected six different target detection methods: LRX (local Reed Xiaoli), CRD (collaborative representation-based detector), LRASR (low-rank and sparse representation), LSMAD (low-rank and sparse matrix decomposition-based Mahalanobis distance method), PTA (prior-based tensor approximation for anomaly detection), and SSIIFD (spectral spatial improved isolation forest detection).

[0131] The test was conducted using spectral images acquired by a snapshot multispectral camera. The 9-channel mosaic snapshot multispectral camera, corresponding to the 600-800nm ​​wavelength band, has the following schematic diagram of its filter array channels: Figure 4 As shown, the acquired image data has a spectral channel resolution of 341*341*9 and a data format of 10-bit RAW. During shooting, the camera was set to a fixed exposure time of 20ms, and the image was saved in 16-bit TIFF format. Referring to relevant work on spectral target detection, the performance of different target detection methods was evaluated and compared using two objective quantitative evaluation metrics: area under the curve (AUC) and processing time.

[0132] Based on the above premises, obtain the spectral image of the region where the large target is located, see [link / reference]. Figure 5 As shown, the six existing target detection methods and the target detection method described in this embodiment are used respectively to detect the target. Figure 5 The spectral image shown was processed, and the results are shown in [reference]. Figure 6 As shown in (a) to (g), the area under the curve (AUC) of the detection results obtained by each detection method is shown in Figures (a to g). Figure 9 As indicated in the line for "major goals," skilled technical personnel can determine [the appropriate objectives]. Figure 6 (a) shows the worst detection result based on the LRX algorithm, which can only detect part of the edge contour of the camouflage target. Figure 6(b) The detection method of the CRD algorithm can detect the complete edge contour of the target, but some areas inside the target are similar to the background pixels. Figure 6 (c) shows that the LRASR algorithm suppresses the background, but there are still regions inside the target that are similar to the background. Figure 6 (d) shows that the LSMAD algorithm can identify the target, but the target and the background are not very different, and there are many false detection areas. Figure 6 The detection results of the PTA algorithm shown in (e) can suppress the background well and can distinguish the target from the background well, but the edge of the target is not clear. Figure 6 The SSIIFD algorithm detection results shown in (f) can also suppress the background to some extent and detect the target area and edge contours, but there are obvious false detection areas. Figure 6 (g) shows the detection results of this embodiment. The target and background are very different. The area where the target is located and the edge contour are very clear. There are no obvious false detection points.

[0133] Further according to Figure 9 The area under the curve (AUC) shown in the "Large Target" row is 0.9963 for the detection method described in this embodiment, which is higher than the detection results of the other six existing methods.

[0134] Furthermore, according to Figure 10 By comparing the image processing times of various detection methods shown in the "Large Target" row, it can be determined that the processing time of the detection method described in this embodiment, 0.0231, is significantly improved compared with existing similar technologies.

[0135] Based on the preceding conditions, if the same small target is detected again, the resulting spectral image containing the small target is as follows: Figure 7 As shown, the six existing target detection methods and the target detection method described in this embodiment are used respectively to detect the target. Figure 7 The spectral image shown was processed, and the results are shown in [reference]. Figure 8 As shown in (a) to (g), the area under the curve (AUC) of the detection results obtained by each detection method is shown in Figures (a to g). Figure 9 As shown in the "Small Goals" row. Technical personnel can determine this. Figure 8 (a) shows the worst detection result based on the LRX algorithm, which can only detect the edge contour of the camouflage target and does not detect the target area. Figure 8 (b) The detection method of the CRD algorithm shown can detect the target area, but the background suppression is poor. Figure 8 Although the LRASR algorithm shown in (c) can detect the target region, there are many similar regions in the background. Figure 8Although the LSMAD algorithm shown in (d) can identify the target area, the distinction between the target and the background area is poor. Figure 8 The PTA algorithm shown in (e) and Figure 8 (f) shows that the SSIIFD algorithm can suppress the background and clearly distinguish the target from the background. However, the SSIIFD algorithm has obvious false detection areas. Figure 8 (g) shows the detection result of this embodiment, where the grayscale difference between the target and the background is large, resulting in good contrast. Further based on... Figure 9 The area under the curve (AUC) shown in the "small target" row indicates that the detection method described in this embodiment achieves 0.9989, which is higher than the detection results of the other six existing methods. Furthermore, according to... Figure 10 The image processing times of various detection methods shown in the "Small Targets" row can be compared to determine that the processing time of the detection method described in this embodiment, 0.0226, is significantly improved compared with existing similar technologies.

[0136] Furthermore, the detection results of the same detection method for large and small targets are compared as follows: The AUC values ​​of different algorithms for large and small targets are... Figure 9 The results are given in [the document]. A value closer to 1 indicates greater effectiveness of the method in detection. The results show that the AUC values ​​for target datasets of different sizes are all above 0.996. The LRX algorithm has the worst AUC value, indicating poor detection performance and robustness. The CRD algorithm also performs poorly in terms of AUC value, with large data fluctuations, indicating poor detection performance and robustness. The LRASR and LSMAD algorithms exhibit large AUC value fluctuations, indicating poor robustness. The PTA and SSIIFD algorithms both have AUC values ​​greater than 0.98 with smaller fluctuations, indicating good detection performance and robustness.

[0137] The spectral target detection method proposed in this embodiment has an AUC value closest to 1, with an AUC value of 0.9963 for large target detection and 0.9989 for small target detection, indicating the most stable performance.

[0138] The processing time of different algorithms for large and small targets is as follows: Figure 10As shown, in both datasets, the SSIIFD algorithm has the longest processing time, exceeding 36 seconds, indicating poor timeliness. This is because the SSIIFD algorithm constructs a large number of iTrees using Gabor features and segmented HSI data during processing, resulting in a long processing time. However, this complexity is traded for better detection results. LRX, as a classic algorithm, has a shorter processing time, only slightly longer than the algorithm presented in this paper, but its results are poor. The CRD algorithm has an average processing time exceeding 4 seconds, indicating low efficiency. The LSMAD algorithm's processing time is only slightly longer than the LRX algorithm, but its performance is an improvement over LRX. The LRASR algorithm has a longer processing time than the LSMAD algorithm, but its robustness is better than LSMAD. The PTA algorithm has the second longest processing time, only shorter than the SSIIFD algorithm, with an average time exceeding 12 seconds, but its detection performance is close to that of the SSIIFD algorithm.

[0139] The spectral target detection method proposed in this embodiment has the shortest processing time, with an average time of only 0.0229s. The processing time for large target detection is 0.0231s, and the processing time for small target detection is 0.0226s. It can process spectral image data from a snapshot spectral camera at a rate of no more than 43 frames per second.

[0140] Implementation Method Twelve: This implementation method compares the spectral target detection method of the present invention with existing similar technologies by conducting target detection experiments on the same large target and small target, and then compares the detection results to prove that the spectral target detection method of the present invention has a faster detection speed and better detection accuracy.

[0141] The spectral target detection method used in this embodiment is based on the spectral target detection method described in Embodiment 1, combined with the optimized steps S1 of Embodiment 5, S2 of Embodiment 7, and S3 of Embodiment 9. The flowchart is as follows: Figure 11 As shown.

[0142] This embodiment provides an example. The six different similar technologies selected in this embodiment are the same as those in the embodiment of embodiment eleven. The detection conditions selected in this embodiment include snapshot multispectral camera acquisition, which are the same as those in the embodiment of embodiment eleven. Therefore, this embodiment compares with the six target detection methods described above:

[0143] like Figure 5 As shown, the six existing target detection methods and the target detection method described in this embodiment are used respectively to detect the target. Figure 5 The spectral image of the region where the large target is located was processed, and the results are shown in [reference]. Figure 6 As shown in (a) to (f) and (h), the area under the curve (AUC) of the detection results obtained by each detection method is shown in Figures 1-2. Figure 9 As shown in the "Large Target" row, those skilled in the art can confirm that the spectral target detection method proposed in this embodiment shows the best results, with a large difference between the target and the background, clear target location and edge contours, and no obvious false detection points.

[0144] Further according to Figure 9 The area under the curve (AUC) shown in the "Large Target" row is 0.9963 for the detection method described in this embodiment, which is higher than the detection results of the other six existing methods.

[0145] Furthermore, according to Figure 10 By comparing the image processing times of various detection methods shown in the "Large Target" row, it can be determined that the processing time of the detection method described in this embodiment, 0.0221, is significantly improved compared with existing similar technologies.

[0146] Based on the preceding conditions, if the same small target is detected again, the resulting spectral image containing the small target is as follows: Figure 7 As shown, the six existing target detection methods and the target detection method described in this embodiment are used respectively to detect the target. Figure 7 The spectral image shown was processed, and the results are shown in [reference]. Figure 8 As shown in (a) to (f) and (h), Figure 8 (h) shows the detection result of this embodiment, where the grayscale difference between the target and the background is large, resulting in good contrast. The area under the curve (AUC) of the detection results obtained by each detection method can be found in [reference needed]. Figure 9 As shown in the "Small Goals" row.

[0147] Further according to Figure 9 The area under the curve (AUC) shown in the "small target" row is 0.9998 for the detection method described in this embodiment, which is higher than the detection results of the other six existing methods.

[0148] Furthermore, according to Figure 10 The image processing times of various detection methods shown in the "Small Targets" row can be compared to determine that the processing time of the detection method described in this embodiment, 0.0237, is significantly improved compared to existing similar technologies.

[0149] Furthermore, the detection results of the same detection method for large and small targets are compared as follows: The AUC values ​​of different algorithms for large and small targets are as follows: Figure 9As shown in the figure, the closer the AUC value is to 1, the more effective the method is in detection. The results show that the AUC value of the spectral target detection method proposed in this embodiment is closest to 1, and the AUC values ​​of different target datasets are all higher than 0.996. The AUC value for large target detection is 0.9963, and the AUC value for small target detection is 0.9998, indicating the most stable performance.

[0150] The processing time of different algorithms for large and small targets is as follows: Figure 10 As shown. The spectral target detection method proposed in this embodiment has the shortest processing time, with an average time of only 0.0229s. The processing time for large target detection is 0.0221s, and the processing time for small target detection is 0.0237s, with an average time of only 0.0229s. It can process spectral image data from a snapshot spectral camera at a rate of no more than 43 frames per second.

Claims

1. A method for detecting spectral targets, characterized in that, Includes the following steps: Step S1: Acquire the spectral image of the region where the target to be detected is located, and perform dimensionality reduction on the spectral image to obtain a three-band pseudo-color image; the spectral image includes spectral images of multiple channels; Step S2: Perform salient region detection on the three-band pseudo-color image to obtain the salient region detection result; Step S3: Perform edge contour detection on the three-band pseudo-color image to obtain the edge contour detection result; Step S4: The salient region detection results and edge contour detection results are weighted and fused to obtain the spectral target detection results; Step S2 is as follows: Step A21: Perform a Fourier transform on the three-band pseudo-color image to obtain a complex matrix, wherein the complex matrix includes an amplitude spectrum and a phase spectrum; Step A22: Perform a logarithmic transformation on the amplitude spectrum to obtain a logarithmic amplitude spectrum; filter the logarithmic amplitude spectrum to obtain a smoothed amplitude spectrum; The filtering method used is either mean filtering or Gaussian filtering. Step A23: Subtract the smoothed amplitude spectrum obtained in step A22 from the amplitude spectrum obtained in step A21 to obtain the spectral residual; Step A24: Combine the phase spectrum and the spectral residual and perform an inverse Fourier transform to obtain the salient region detection result.

2. The spectral target detection method according to claim 1, characterized in that, Step S3 is as follows: Step A31: Perform the Canny edge detection algorithm on the three-band pseudo-color image to obtain the edge results of the three channels in the three-band pseudo-color image; accumulate and perform binarized threshold segmentation on the image edge results of the three channels to obtain preliminary edge contour results; Step A32: Perform image opening operation on the obtained preliminary edge contour results to obtain the edge contour detection results.

3. A method for detecting spectral targets, characterized in that, Includes the following steps: Step S1: Acquire the spectral image of the region where the target to be detected is located, and perform dimensionality reduction on the spectral image to obtain a three-band pseudo-color image; the spectral image includes spectral images of multiple channels; Step S2: Perform salient region detection on the three-band pseudo-color image to obtain the salient region detection result; Step S3: Perform edge contour detection on the three-band pseudo-color image to obtain the edge contour detection result; Step S4: The salient region detection results and edge contour detection results are weighted and fused to obtain the spectral target detection results; Step S2 is as follows: Step B21: Convert the three-band pseudo-color image to the XYZ color space to obtain the XYZ pseudo-color image; Step B22: Convert the XYZ pseudo-color image to the LAB color space to obtain a LAB color image; Step B23: Based on the LAB color image, obtain the LAB color vector of each pixel and the average LAB color vector of all pixels; Step B24: The salient region detection result is obtained based on the Euclidean distance between the LAB color vector of each pixel and the average LAB color vector of the image.

4. The spectral target detection method according to claim 3, characterized in that, Step S3 is as follows: Step B31: The gradient magnitudes of the three channels in the three-band pseudo-color image are obtained by applying the Sobel algorithm to the three-band pseudo-color image. Step B32: Accumulate the gradient magnitudes of the three channels and perform binarization thresholding to obtain preliminary edge contour results; The gradient magnitude includes the horizontal gradient magnitude and the vertical gradient magnitude; Step B33: Perform image opening operation on the obtained preliminary edge contour results to obtain the edge contour detection results.

5. The spectral target detection method according to claim 1 or 3, characterized in that, Step S1 includes: Step S11: Select the spectral images of three channels from the spectral image; Step S12: Based on the spectral images of the three channels, obtain a three-band pseudo-color image.

6. The spectral target detection method according to claim 5, characterized in that, Step S11 includes: Step S111: For the spectral image of each channel in the spectral image, obtain its corresponding spectral average response image; Step S112: Based on the average response images of all spectra, obtain the spectral difference score of the spectral image for each channel; Step S113: Sort the spectral images of each channel according to the spectral difference score, and select the spectral images of the three channels with the highest spectral difference scores.

7. The spectral target detection method according to claim 6, characterized in that, The spectral difference score of the spectral image of each channel in step S112. pass Obtained, wherein the spectral image contains * 1 pixel; For the first The coordinates in the spectral channel image of each channel are The grayscale value of the pixel; The coordinates in the spectral average response image are The grayscale value of the pixel.

8. The spectral target detection method according to claim 6, characterized in that, The spectral difference fraction mentioned in step S112 pass get.

9. The spectral target detection method according to claim 1 or 3, characterized in that, In step S4, the weights for weighted fusion are all set to 0.5.

Images