A hyperspectral image target intelligent detection method based on deep learning

Abstract

The application discloses a hyperspectral image target intelligent detection method based on deep learning and belongs to the field of remote sensing image processing. The method comprises the following steps: selecting a background training set with large difference and representativeness between each other through a coarse detection and de-correlated background pixel selection strategy; generating sufficient target training sets by using a target spectrum band random replacement target training sample data enhancement strategy; mixing and disordering the background and target training sets, and inputting the background and target training sets into upper and lower branches of a multi-depth-multi-branch target detection network respectively for training; inputting a to-be-detected pixel and a prior target into the upper and lower branches of the trained network respectively for target detection, and outputting the network to form a final detection result. By using the application, target pixels in a hyperspectral image can be detected and highlighted, and background pixels can be effectively suppressed, so that a detection result with high detection precision, high background suppression degree and high target-background separation degree is obtained.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral remote sensing image processing technology, specifically relating to a deep learning-based intelligent target detection method for hyperspectral images. Background Technology

[0002] Hyperspectral images are three-dimensional images with rich spectral bands acquired by spectral imagers. Their spectral resolution reaches the nanometer level, providing detailed, reliable, and abundant information about ground features. Therefore, they are widely used in fields such as ground feature classification, target detection, and anomaly detection. Hyperspectral target detection is a pixel-by-pixel detection method that relies on limited prior target spectra and utilizes the difference in spectral information between the background and the target feature. It has been extensively studied in recent decades and plays a crucial role in fields such as biomedicine, ground feature observation, military reconnaissance, and mineral exploration. Hyperspectral images used for hyperspectral target detection typically contain only positive and negative labels. Furthermore, the datasets contain a very small number of sparsely distributed target pixels. Therefore, how to quickly and accurately locate the target of interest in a large amount of complex background while ensuring sufficient separation between the background and the target is the main challenge of hyperspectral target detection algorithms.

[0003] Over the past few decades, numerous scholars have proposed various detection algorithms for hyperspectral target detection. Traditional hyperspectral target detection algorithms mainly include those based on probabilistic statistical models, original spatial models, and subspace projection models. Most traditional detectors are based on linear mixture models where a pixel is composed of multiple ground features, and are designed based on the assumption of a multivariate normal distribution of the target and background. This approach only achieves good detection results under ideal conditions. However, real hyperspectral images typically exhibit spectral variability and nonlinear correlations between spectra, causing traditional hyperspectral target detectors to fall short of their intended detection levels and creating difficulties for practical applications. To better adapt to target detection in real-world scenarios, scholars have proposed many machine learning-based hyperspectral target detection algorithms, which can be mainly categorized into those based on sparse representation and cooperative representation, low-rank decomposition, tensor decomposition, and kernel functions. Machine learning-based hyperspectral target detectors are designed based on the characteristics of hyperspectral images and can achieve high detection accuracy in practical applications. However, they often require manual feature selection and extraction, resulting in a lack of feature expressiveness.

[0004] In recent years, end-to-end deep learning, which requires no expert experience, has provided new insights for hyperspectral object detection due to its powerful ability to automatically mine inherent patterns in data and extract deep features. Some deep learning-based hyperspectral object detection algorithms have achieved good detection performance in complex real-world scenes, but several issues remain. First, selecting a representative background training set with significant differences from a large number of complex background pixels is a challenge for hyperspectral object detection. Furthermore, constructing reasonable target training samples that are categorically balanced with the background training set using limited prior targets is one of the biggest challenges facing deep learning-based hyperspectral object detection. Finally, and equally importantly, how to leverage deep networks to better suppress the background while detecting and highlighting targets, thereby improving the separation between targets and background and optimizing the visual effect of detection, is a meaningful and challenging problem. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing deep learning-based hyperspectral target detection technologies and provide a deep learning-based intelligent target detection method for hyperspectral images with high detection accuracy and high target prominence-background separation.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] S1: Select hyperspectral image Training set of representative backgrounds that are unrelated to each other Where d and N are the number of spectral bands of pixels in the hyperspectral image and the total number of pixels contained in the hyperspectral image, respectively;

[0008] S2: For a priori target A sufficient target training set is generated by randomly replacing bands.

[0009] S3: Transfer background training samples and target training samples Randomly shuffled, and compared with a prior target pixel x t The inputs are fed into the upper and lower branches of the multi-depth-multi-branch object detection network for training to obtain the trained model;

[0010] As a preferred embodiment of the present invention: the multi-depth-multi-branch object detection network includes upper and lower branches. During training, the upper branch receives a shuffled background and target training set as input. and The lower branch input is a prior target pixel x. t The multi-depth-multi-branch object detection network includes a multi-depth feature extraction module (MDFE), a local-global parallel feature extraction module (Conv-GRU), and multi-branch feature fusion.

[0011] S4: Input the pixel to be tested in the hyperspectral image and a prior target into the upper and lower branches of the trained multi-depth-multi-branch target detection network respectively to obtain the predicted value corresponding to each pixel to be tested;

[0012] S5: Use the obtained predicted value as the target detection result of the pixel to be tested, and obtain the target detection result of the entire hyperspectral image.

[0013] As a preferred technical solution of the present invention: Step S1 uses a background sample selection strategy based on spectral angle mapping SAM coarse detection and decorrelation, that is, using SAM and Kullback-Leibler (KL) divergence to select a background training set that is uncorrelated and representative in the hyperspectral image. The specific implementation method is as follows:

[0014] A1: Calculate the relationship between each pixel in the hyperspectral image X and the prior target pixel x based on the SAM (Spectral Angle Mapping) method. t The spectral angle is expressed as

[0015]

[0016] A2: Sort the spectral angles in Θ from largest to smallest, with larger spectral angles indicating higher priority. Then, select the first N × 0.99 original pixels as candidate background pixels, denoted as...

[0017]

[0018] A3: Remove The highest priority pixel in the current dataset is added to the background training set. In the middle, and delete The corresponding pixel in the middle;

[0019] A4: Will The highest priority pixel and background training set in the current dataset The KL divergence value δ of each pixel is calculated sequentially, and each time δ is obtained, it is compared with the threshold ε.

[0020] A5: Delete if there is a case where δ is less than the set threshold ε. If the pixel with the highest priority is selected, proceed to step A4; otherwise, after all δ comparisons are completed, proceed to step A3, and so on.

[0021] A6: Select the background training set Each pixel is assigned a negative sample label of 0, and calculation is performed. The total number of pixels in the array is denoted as N. s .

[0022] As a preferred technical solution of the present invention: in step S2, a target spectrum random replacement strategy is used to replace a priori target x t = [l1, l2, ..., l d ] T A new target sample is obtained by replacing the spectral band values ​​at a random number of random locations with the average of all the band values ​​to be replaced. This process is repeated to generate a target training set that is balanced with the background training set. The specific implementation steps are as follows:

[0023] B1: Initialize the target training set Initialize k = 1;

[0024] B2: Randomly generate the number of band replacements num∈[1,d], and initialize the k-th target training sample as... Initialize the sum of values ​​for the bands to be replaced to 0;

[0025] B3: Randomly generate band replacement position pos h ∈[1, d] (h = 1, 2, ..., num), and calculate the sum of the band values ​​at the corresponding positions. and mean

[0026] B4: Replace the band values ​​corresponding to all replacement band positions with the mean value to generate new target training samples. The process expression is:

[0027]

[0028] B5: Will Add it to the target training set, i.e.

[0029] B6: When k≤N s If k = k + 1, proceed to step B2; otherwise, obtain the generated target training set. Each pixel is then assigned the positive sample label 1.

[0030] As a preferred technical solution of the present invention: In the multi-depth feature extraction module (MDFE) of the multi-depth-multi-branch target detection network described in step S3, a one-dimensional convolution with a kernel size of 1, a stride of 1, and a number of kernels of 16 is used to reduce the original number of channels to 1 / 4 so that the dimension is the same when the features are fused in the future; four one-dimensional convolutions with kernel sizes of 3, 5, 7, and 9, a stride of 1, and a number of kernels of 16 are used to extract multi-size spectral local feature information. Each one-dimensional convolution has two outputs: one is to retain the features after its own convolution for subsequent feature channel concatenation, and the other is to input to the next convolution of different scales to obtain features of different depths; then the features obtained from the four convolutions of different sizes are concatenated using channels and passed through a one-dimensional convolution with a kernel size of 3, a stride of 2, and a number of kernels of 64 to reduce the resolution, filter redundant feature information, and extract key local features of the spectrum.

[0031] As a preferred technical solution of the present invention: In the local-global parallel feature extraction module Conv-GRU of the multi-depth-multi-branch target detection network described in step S3, two 1D convolutions with a kernel size of 3, strides of 2 and 1 respectively, and a kernel number of 64 are used for spectral local feature extraction, and a gated recurrent unit (GRU) is used for spectral global feature extraction. The local convolution and the GRU are connected in parallel to form Conv-GRU.

[0032] As a preferred technical solution of the present invention: the multi-branch feature fusion of the multi-depth-multi-branch target detection network in step S3 uses three operations: element-wise addition, element-wise multiplication, and element-wise subtraction, to obtain feature maps f from the background / target training set and the prior target respectively through the multi-depth feature extraction module MDFE. D and f P The fusion yields three different relation feature maps, represented as follows:

[0033] f + =f D +f P , ( (representing element-wise multiplication) and f - =f D -f P ,

[0034] To obtain sufficient correlation features between the training set and the prior target; then, f + f × f - f D and f PFive different features are input into the local-global parallel feature extraction module Conv-GRU to obtain 10 different local-global feature information. After channel splicing and fusion, the feature discrimination capability of the network is increased. Furthermore, three one-dimensional convolutions with 3 kernels, strides of 2 and 1, and a total of 64 kernels are used to extract key features. Fully connected layers and sigmoid activation function are used to make the detection output value between 0 and 1.

[0035] As a preferred embodiment of the present invention: the multi-depth-multi-branch object detection network described in step S3 uses a binary cross-entropy loss function to optimize training, the expression of which is:

[0036]

[0037] Where B represents the batch size, f i Represents the final output of the network, y i Represents a positive or negative category label of 1 or 0.

[0038] The present invention provides a deep learning-based intelligent target detection method for hyperspectral images. Compared with existing technologies, the advantages of the present invention are as follows:

[0039] (1) For the background training set, this invention proposes a background sample selection strategy based on spectral angle mapping SAM coarse detection and decorrelation, which selects background pixels with low correlation between them and representative background pixels, providing reasonable background training samples for the subsequent training stage of the target detection network.

[0040] (2) For the target training set, this invention proposes a target training sample data augmentation method with random target spectrum replacement. Sufficient target training samples are generated by repeatedly replacing the band values ​​of a prior target at random numbers and random positions with the mean of all bands to be replaced. Specifically, the random target spectrum replacement strategy only requires one prior target and no other data. Furthermore, it can simulate spectral changes in real-world scenes, increasing sample diversity and improving the generalization performance of subsequent networks.

[0041] (3) To highlight the target and suppress the background, this invention proposes a multi-depth-multi-branch target detection network. First, a multi-depth feature extraction module is used to learn the multi-depth spectral local features of the sample and the prior target. Then, by using the cooperation of convolution and gated recurrent units (GRUs), the spectral local-global features of the sample and the prior target itself, as well as the interrelationships obtained through different fusion strategies, are fully explored. This enables the network to automatically identify and highlight target pixels that are consistent with the prior target and suppress background pixels that are inconsistent with it, thereby improving the final detection performance. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating the method provided for an example of the present invention;

[0044] Figure 2 A schematic diagram of the background sample selection process provided for an example of the present invention;

[0045] Figure 3 A schematic diagram illustrating the process of generating target samples for an example of the present invention;

[0046] Figure 4 A structural diagram of a multi-depth-multi-branch target detection network provided for an example of the present invention;

[0047] Figure 5 The images shown are real ground cover maps and detection results from the Urban hyperspectral dataset used in the simulation experiment of this invention.

[0048] Figure 6 The images shown are real ground cover maps and detection results from the Chikusei hyperspectral dataset used in the simulation experiment of this invention.

[0049] Figure 7 This is the ROC curve of the detection results of the Urban hyperspectral dataset in the simulation experiment of this invention;

[0050] Figure 8 This is the ROC curve of the detection results of the Chikusei hyperspectral dataset in the simulation experiment of this invention;

[0051] Figure 9 This is a separation mapping diagram of the detection results of the Urban hyperspectral dataset in the simulation experiment of this invention;

[0052] Figure 10 This is a separation mapping diagram of the detection results of the Chikusei hyperspectral dataset in the simulation experiment of this invention. Detailed Implementation

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0054] This invention discloses a deep learning-based intelligent target detection method for hyperspectral images, such as... Figure 1As shown, it specifically includes three parts: background training set selection, target training set generation, and a multi-depth, multi-branch object detection network. Figure 2 As shown, in the selection of the background training set, a background sample selection strategy based on spectral angle mapping (SAM) coarse detection and decorrelation is used to select representative background pixels with low correlation between them, and these pixels are combined with the negative label 0 to form the background training set. Figure 3 As shown, in the generation of the target training set, a data augmentation method using random replacement of the target spectrum is used to generate sufficient target training samples, which, together with the positive label 1, constitute the target training set. Figure 4 The diagram shows a multi-depth, multi-branch network structure. During network training, the background and target training sets are randomly shuffled and input into the upper branch of the network, while a prior target is input into the lower branch, resulting in a trained model. During the network testing phase, i.e., the target detection phase, all pixels to be tested in the hyperspectral image and a prior target are input into the upper and lower branches of the trained network, respectively. After passing through the trained network, the detection value is obtained as the detection result of the pixel to be tested, and finally, the hyperspectral target detection result is constructed.

[0055] In this embodiment, the prior target used is obtained by randomly selecting five target pixels from the hyperspectral image based on the ground truth map and averaging them.

[0056] In the embodiment, the matrix corresponding to the hyperspectral image and a priori target pixel spectral vector (Where d and N are the number of spectral bands per pixel and the total number of pixels, respectively) Perform the following operations:

[0057] S1: A background sample selection strategy based on SAM coarse detection and decorrelation using spectral angle mapping, employing SAM and KL divergence to select hyperspectral images. The training set consists of representative background pixels that are not related to each other, and the negative label 0.

[0058] In this embodiment, the algorithm flow of step S1 is as follows: Figure 2 As shown, the specific operation includes the following steps:

[0059] A1: Calculate the relationship between each pixel in the hyperspectral image X and the prior target pixel x based on the SAM (Spectral Angle Mapping) method. t The spectral angle is expressed as

[0060]

[0061] A2: Sort the spectral angles in Θ from largest to smallest, with larger spectral angles indicating higher priority. Then, select the first N × 0.99 original pixels as candidate background pixels, denoted as...

[0062]

[0063] A larger SAM (Spectral Angle of Interest) value indicates a less similarity between the pixel and the prior target pixel, meaning the pixel is more likely to be a background pixel and thus has a higher priority. Furthermore, in hyperspectral images used for hyperspectral target detection, target pixels typically account for less than 1% of the total pixels. Therefore, extracting the top 99% of pixels with larger SAM values ​​effectively prevents subsequent selection of target pixels, thus avoiding an impure background training set.

[0064] A3: Remove The highest priority pixel in the current dataset is added to the background training set. In the middle, and delete The corresponding pixel in the middle;

[0065] A4: Will The highest priority pixel and background training set in the current dataset The KL divergence value δ of each pixel is calculated sequentially, and each time δ is obtained, it is compared with the threshold ε.

[0066] A5: Delete if there is a case where δ is less than the set threshold ε. If the pixel with the highest priority is selected, proceed to step A4; otherwise, after all δ comparisons are completed, proceed to step A3, and so on.

[0067] KL divergence measures the degree of difference between two pixels; the smaller the difference, the smaller the KL divergence. Therefore, by using SAM and KL divergence as metrics, the selected background samples can be representative and have significant differences from each other, providing more reasonable training samples for network training. A preferred threshold ε is set to 0.002, but this should be adjusted based on the specific dataset used in implementation.

[0068] A6: Select the background training set Each pixel is assigned a negative sample label of 0, and calculation is performed. The total number of pixels in the array is denoted as N. s .

[0069] Spectral angle mapping SAM and Kullback-Leibler divergence are well-known techniques in this field and will not be elaborated upon here.

[0070] S2: For a priori target A sufficient target training set is generated by randomly replacing bands.

[0071] In this embodiment, step S2 utilizes a target spectrum random replacement strategy to replace a priori target x t = [l1, l2, ..., ld ] T A new target sample is obtained by replacing the spectral band values ​​at a random number of random locations with the average of all the band values ​​to be replaced. This process is repeated to generate a target training set that is balanced with the background training set. To ensure a balance between positive and negative samples, the number of target samples generated should be the same as the number of background training samples, which is N. s , Figure 3 The diagram shows the algorithm flow for the target band random replacement strategy. The specific operations include the following steps:

[0072] B1: Initialize the target training set Initialize k = 1;

[0073] B2: Randomly generate the number of band replacements num∈[1,d], and initialize the k-th target training sample as... Initialize the sum of values ​​for the bands to be replaced to 0;

[0074] B3: Randomly generate band replacement position pos h ∈[1, d] (h = 1, 2, ..., num), and calculate the sum of the band values ​​at the corresponding positions. and mean

[0075] B4: Replace the band values ​​corresponding to all replacement band positions with the mean value to generate new target training samples. The process expression is:

[0076]

[0077] B5: Will Add it to the target training set, i.e.

[0078] B6: When k≤N s If k = k + 1, proceed to step B2; otherwise, obtain the generated target training set. Each pixel is then assigned the positive sample label 1.

[0079] The target spectral random replacement strategy can simulate the spectral changes of real-world scenes, increasing sample diversity. Furthermore, this process only requires one prior target, reducing resource costs and ensuring that the generated target training samples do not mix with background pixel information. This effectively avoids false detections in subsequent networks and improves the generalization performance of those networks.

[0080] S3: Transfer background training samples and target training samples Randomly shuffled, and compared with a prior target pixel x tThe inputs are fed into the upper and lower branches of the multi-depth-multi-branch object detection network for training, resulting in a trained model.

[0081] In this embodiment, the structure diagram of the multi-depth-multi-branch target detection network described in step S3 is as follows: Figure 4 As shown, it includes two branches (upper and lower), a multi-depth feature extraction module (MDFE), a local-global parallel feature extraction module (Conv-GRU), and a multi-branch feature fusion module. The specific structure and operation of each module are as follows:

[0082] In the Multi-Depth Feature Extraction (MDFE) module of the Multi-Depth-Multi-Branch Target Detection Network, a one-dimensional convolution with a kernel size of 1, a stride of 1, and 16 kernels is used to reduce the original number of channels to 1 / 4 so that the dimension remains the same during subsequent feature fusion. Four one-dimensional convolutions with kernel sizes of 3, 5, 7, and 9, a stride of 1, and 16 kernels are used to extract multi-scale spectral local feature information. Each one-dimensional convolution has two outputs: one is to retain the features after its own convolution for subsequent feature channel concatenation, and the other is to input to the next convolution at a different scale to obtain features at different depths. Then, the features obtained from the four convolutions of different sizes are concatenated using channels and passed through a one-dimensional convolution with a kernel size of 3, a stride of 2, and 64 kernels to reduce the resolution, filter redundant feature information, and extract key local spectral features.

[0083] In the Conv-GRU local-global parallel feature extraction module of the multi-depth-multi-branch target detection network, two 1D convolutions with a kernel size of 3, strides of 2 and 1 respectively, and a total of 64 kernels are used for spectral local feature extraction. A gated recurrent unit (GRU) is used for extracting spectral global features. The local convolutions and the GRU are connected in parallel to form the Conv-GRU. The gated recurrent unit (GRU) is a well-known technology in this field and will not be described in detail here.

[0084] The multi-branch feature fusion of the multi-depth-multi-branch object detection network uses three operations: element-wise addition, element-wise multiplication, and element-wise subtraction. The background / target training set and the prior target are processed by the multi-depth feature extraction module (MDFE) to obtain feature maps f. D and f P The fusion yields three different relation feature maps, represented as follows:

[0085] f + =f D +f P , ( (representing element-wise multiplication) and f - =f D -f P ,

[0086] To obtain sufficient correlation features between the training set and the prior target; then, f + f × f - f D and f P Five different features are input into the local-global parallel feature extraction module Conv-GRU to obtain 10 different local-global feature information. After channel splicing and fusion, the feature discrimination capability of the network is increased. Furthermore, three one-dimensional convolutions with 3 kernels, strides of 2 and 1, and a total of 64 kernels are used to extract key features. Fully connected layers and sigmoid activation function are used to make the detection output value between 0 and 1.

[0087] The multi-depth-multi-branch object detection network described above uses a binary cross-entropy loss function to optimize training, the expression of which is:

[0088]

[0089] Where B represents the batch size, f i Represents the final output of the network, y i The label represents the positive or negative category, 1 or 0. The binary cross-entropy loss function is a well-known technique in this field.

[0090] The preferred recommended batch size for network training is 512, and the learning rate is 10. -4 The number of training epochs is set to 100.

[0091] S4: Input the pixel to be tested and a prior target from the hyperspectral image into the upper and lower branches of the trained multi-depth-multi-branch target detection network respectively to obtain the predicted value corresponding to each pixel to be tested.

[0092] In this embodiment, step S4 specifically involves inputting each pixel to be measured in the hyperspectral image X and a priori target pixel spectral vector x. t Input the upper and lower branches of the multi-depth multi-branch object detection network, use the trained network model to automatically determine the probability that the pixel to be tested is an object pixel, and output the corresponding prediction value.

[0093] S5: Use the obtained predicted value as the target detection result of the pixel to be tested, and obtain the target detection result of the entire hyperspectral image.

[0094] In this embodiment, step S5 specifically involves the following steps: the multi-depth-multi-branch target detection network outputs predicted values ​​for N pixels to be tested. The higher the predicted value, the higher the probability that the pixel is a target pixel. The closer the predicted value is to 0, the higher the probability that the pixel is a background pixel. Finally, the target detection result of the hyperspectral image is obtained.

[0095] In practical implementation, the process can be automated using software. The device for running the process should also be within the scope of protection of this invention. The deep learning experiments in the embodiments were conducted using Python 3.9.12, Conda 4.12.0, and PyTorch 1.2.0; all others were implemented using MATLAB R2019a.

[0096] Based on the above scheme, the following comparative experiments are used to verify the beneficial effects of the method of the present invention.

[0097] In this embodiment, the public hyperspectral data used are the Urban hyperspectral dataset and the Chikusei hyperspectral dataset. The Urban dataset comes from the urban dataset in the "Airport-Beach-City" dataset. It is a photograph taken by AVIRIS on August 29, 2010, of the Texas coast, with a spatial sampling distance of 17.2 meters. After removing low-quality spectral bands, it contains 207 spectral bands, with a two-dimensional image size of 100 pixels × 100 pixels. Buildings consisting of 88 pixels are selected as the detection targets of interest. The Chikusei dataset comes from a rural and urban border area in Chikusai City, Ibaraki Prefecture, Japan, taken by Headwall Hyperspec-VNIR-C on July 29, 2014. It has a spatial sampling distance of 2.5 meters, 128 spectral bands, and a two-dimensional image size of 150 pixels × 150 pixels. The target is a vehicle on a road consisting of 15 pixels.

[0098] In this embodiment, four comparative algorithms are used to compare and analyze the method of the present invention. The comparative algorithms are: the classic traditional detection algorithms: Adaptive Cosine / Consistent Estimation Algorithm (ACE) (Method 1) and Constrained Energy Minimization Detector (CEM) (Method 2); the machine learning-based detection algorithm: CSCR (Method 3), a target detection algorithm based on a combination of sparse representation and cooperative representation; and the deep learning-based detection algorithm: TSCND (Method 4), a dual-branch convolutional neural network hyperspectral target detection algorithm. The method of the present invention is exemplified by a specific implementation. The above comparative algorithms are well-known technologies in this technical field and will not be described in detail here.

[0099] In this embodiment, the Receiver Operational Characteristic (ROC) curve, AUC value (area under the ROC curve), and separation map are used as evaluation metrics. Specifically, the closer the ROC curve is to the top left corner of the map, the larger the corresponding AUC value, indicating better detection performance of the algorithm. The separation map can be used to evaluate the degree of separation between the background and target classes in the detection results, and it can also be used to observe the data distribution, verifying the algorithm's suppression of background pixels.

[0100] Table 1. AUC values ​​of the comparative experiments

[0101] AUC value Urban Dataset Chikusei dataset Method 1 ACE 0.95330 0.98377 Method 2 CEM 0.92359 0.99307 Method 3 CSCR 0.99470 0.99361 Method 4TSCNTD 0.99570 0.94408 Method of the present invention 0.99932 0.99664

[0102] As shown in Table 1, compared with other detection algorithms, the method of this invention achieved the highest AUC values ​​for both datasets.

[0103] like Figure 5-9 As shown, Figure 5 and Figure 6 The images show the detection results of different detection algorithms on the Urban and Chikusei datasets, respectively. It can be seen that the method of this invention can basically locate the target position correctly, highlight the target pixels well, and suppress background pixels to a very high degree. Figure 7 and Figure 8 The figures show the ROC curves for the detection results on the Urban and Chikusei datasets, respectively. For the Urban dataset, the ROC curve corresponding to the method of this invention is closest to the top left corner, and its detection probability reaches 1 earliest as the false alarm probability increases. For the Chikusei dataset, although the ROC curve corresponding to ACE of method 3 is initially at the top, its detection probability reaches 1 last, while the ROC curve corresponding to the method of this invention is above other methods, showing the best overall performance. All of the above demonstrates that the method of this invention has higher detection accuracy. Figure 9 and Figure 10 The separation mapping diagrams of the detection results on the Urban and Chikusei datasets shown indicate that the background bins corresponding to the method of this invention are almost completely close to the statistical value of 0, indicating that it has the strongest ability to suppress the background. The larger distance between the background and the target bins indicates that the target-background separation ability of the algorithm of this invention is better. The experimental results show that, compared with other comparative experiments, the target detection results of the method of this invention achieve the best detection accuracy, background suppression degree, and target-background separation effect.

[0104] It should be understood that any undescribed parts of this invention are the same as or implemented using existing technology.

[0105] It should be understood that the present invention is not limited to the above-described embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and all such changes fall within the protection scope of the present invention. The scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for intelligent target detection in hyperspectral images based on deep learning, characterized in that, Includes the following steps: S1: Select hyperspectral image The training set D of unrelated representative backgrounds b , where d and N are the number of spectral bands of pixels in the hyperspectral image and the total number of pixels contained in the hyperspectral image, respectively; S2: For a priori target A sufficient target training set D is generated by randomly replacing bands. t , that is, a priori target x t = [l1, l2, ..., l d ] T A new target sample is obtained by replacing the spectral band values ​​at a random number of random locations with the average of all the band values ​​to be replaced. This process is repeated to generate a target training set D that is balanced with the background training set. t The specific implementation steps are as follows: B1: Initialize the target training set Initialize k = 1; B2: Randomly generate the number of band replacements num∈[1,d], and initialize the k-th target training sample as... Initialize the sum of values ​​for the bands to be replaced to 0; B3: Randomly generate band replacement position pos h ∈[1, d] (h = 1, 2, ..., num), and calculate the sum of the band values ​​at the corresponding positions. and mean B4: Replace the band values ​​corresponding to all replacement band positions with the mean value to generate new target training samples. This process is represented as B5: Will Add it to the target training set, i.e. B6: When k≤N s If k = k + 1, jump to step B2; otherwise, obtain the generated target training set D. t And assign each pixel a positive sample label 1; S3: Transfer the background training samples D b and target training sample D t Randomly shuffled, and compared with a prior target pixel x t The inputs are fed into the upper and lower branches of the multi-depth-multi-branch object detection network for training to obtain the trained model; The multi-depth-multi-branch object detection network includes upper and lower branches; during training, the upper branch is input with a shuffled background and target training set D. b and D t The lower branch input is a prior target pixel x. t The multi-depth-multi-branch object detection network includes a multi-depth feature extraction module (MDFE), a local-global parallel feature extraction module (Conv-GRU), and multi-branch feature fusion. S4: Input the pixel to be tested in the hyperspectral image and a prior target into the upper and lower branches of the trained multi-depth-multi-branch target detection network respectively to obtain the predicted value corresponding to each pixel to be tested; S5: Use the obtained predicted value as the target detection result of the pixel to be tested, and obtain the target detection result of the entire hyperspectral image.

2. The method for intelligent target detection in hyperspectral images based on deep learning according to claim 1, characterized in that: Step S1 employs a background sample selection strategy based on spectral angle mapping using SAM coarse detection and decorrelation, specifically using SAM and Kullback-Leibler (KL) divergence to select an uncorrelated and representative background training set D from the hyperspectral image. b The specific implementation method is as follows: A1: Calculate the relationship between each pixel in the hyperspectral image X and the prior target pixel x based on the SAM (Spectral Angle Mapping) method. t The spectral angle is expressed as A2: Sort the spectral angles in Θ from largest to smallest, with larger spectral angles indicating higher priority. Then, select the first N × 0.99 original pixels as candidate background pixels, denoted as... A3: Extract the highest priority pixel from D and add it to the background training set D. b In the middle, delete the corresponding pixel in D; A4: Combine the highest priority pixel in D with the background training set D. b The KL divergence value δ of each pixel is calculated sequentially, and each time δ is obtained, it is compared with the threshold ε. A5: If any δ is less than the set threshold ε, delete the highest priority pixel in D and proceed to step A4; otherwise, after all δ comparisons are completed, proceed to step A3, and so on. A6: Select the background training set D b Each pixel is assigned a negative sample label of 0, and D is calculated. b The total number of pixels in the array is denoted as N. s .

3. The method for intelligent target detection in hyperspectral images based on deep learning according to claim 1, characterized in that: In the Multi-Depth Feature Extraction (MDFE) module of the multi-depth-multi-branch object detection network described in step S3, a one-dimensional convolution with a kernel size of 1, a stride of 1, and 16 kernels is used to reduce the original number of channels to 1 / 4 so that the dimension remains the same during subsequent feature fusion. Four one-dimensional convolutions with kernel sizes of 3, 5, 7, and 9, a stride of 1, and 16 kernels are used to extract multi-scale spectral local feature information. Each one-dimensional convolution has two outputs: one is to retain the features after its own convolution for subsequent feature channel concatenation, and the other is to input to the next convolution at a different scale to obtain features at different depths. Then, the features obtained from the four convolutions of different sizes are concatenated using channels and passed through a one-dimensional convolution with a kernel size of 3, a stride of 2, and 64 kernels to reduce the resolution, filter redundant feature information, and extract key local spectral features.

4. The method for intelligent target detection in hyperspectral images based on deep learning according to claim 1, characterized in that: In the local-global parallel feature extraction module Conv-GRU of the multi-depth-multi-branch target detection network described in step S3, two 1D convolutions with a kernel size of 3, strides of 2 and 1 respectively, and a kernel count of 64 are used for spectral local feature extraction, and a gated recurrent unit (GRU) is used for spectral global feature extraction. The local convolutions and GRUs are connected in parallel to form Conv-GRU.

5. The method for intelligent target detection in hyperspectral images based on deep learning according to claim 1, characterized in that: The multi-branch feature fusion of the multi-depth-multi-branch object detection network described in step S3 uses element-wise addition, element-wise multiplication, and element-wise subtraction to extract feature maps f from the background / target training set and prior targets obtained by the multi-depth feature extraction module MDFE. D and f P The fusion yields three different relation feature maps, denoted as f. + =f D +f P , ( (representing element-wise multiplication) and f - =f D -f P , To obtain sufficient correlation features between the training set and the prior target; then, f + f × f - f D and f P Five different features are input into the local-global parallel feature extraction module Conv-GRU to obtain 10 different local-global feature information. After channel splicing and fusion, the feature discrimination capability of the network is increased. Furthermore, three one-dimensional convolutions with 3 kernels, strides of 2 and 1, and a total of 64 kernels are used to extract key features. Fully connected layers and sigmoid activation function are used to make the detection output value between 0 and 1.

6. The intelligent target detection method for hyperspectral images based on deep learning according to claim 1, characterized in that: The multi-depth-multi-branch object detection network described in step S3 uses a binary cross-entropy loss function to optimize training, the expression of which is: Where B represents the batch size, f i Represents the final output of the network, y i Represents a positive or negative category label of 1 or 0.

Images