Hyperspectral target detection method and system based on depth separable convolution
By combining a depthwise separable convolutional module and a U-shaped Transformer, the problem of insufficient computational resources in hyperspectral target detection is solved, and efficient feature extraction and target detection are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE SECOND CONSTR OF CHINA CONSTR EIGHTH ENG DIV
- Filing Date
- 2026-01-15
- Publication Date
- 2026-04-24
AI Technical Summary
Existing hyperspectral target detection methods have insufficient utilization of computational resources, especially when processing large-scale data, which requires a large amount of computation and memory. Furthermore, supervised detection methods are costly, and self-supervised and unsupervised detection methods have limited effectiveness in specific scenarios.
A depthwise separable convolution module is used to process hyperspectral images. By combining a U-shaped Transformer and non-negative constraints, standard convolutions are decomposed through depthwise convolution and pointwise convolution, reducing the number of parameters and improving feature extraction capabilities, thereby enhancing the model's generalization ability.
It significantly reduces computational load and memory usage, improves model computational efficiency, enhances feature representation capabilities, and achieves more efficient hyperspectral target detection.
Smart Images
Figure CN121921622A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of hyperspectral image detection, specifically, it relates to a hyperspectral target detection method and system based on depthwise separable convolution. Background Technology
[0002] Hyperspectral imagery (HSI) is a type of image data with unique properties and significant application value. From an imaging principle perspective, it simultaneously images a target area using an imaging spectrometer across the ultraviolet, visible, near-infrared, and mid-infrared regions of the electromagnetic spectrum, employing dozens to hundreds of continuous and subdivided spectral bands. Compared to traditional multispectral images, hyperspectral images offer a significantly richer information content. In traditional multispectral images, each pixel is typically represented by only a few discrete spectral bands, while in hyperspectral images, each pixel is a spectral curve containing hundreds of bands, typically covering the visible to near-infrared or short-wave infrared regions. This allows hyperspectral images to capture more subtle spectral differences in ground features. In terms of applications, hyperspectral images have important uses in various fields such as resource exploration, crop growth monitoring, hazardous substance detection, disaster assessment, and military reconnaissance.
[0003] In the field of hyperspectral target detection, there are three main detection methods: supervised detection, self-supervised detection, and unsupervised detection. Supervised detection methods rely on labeled datasets and learn the spectral features of target pixels to identify and classify targets in images. However, this method may be limited when the distribution of target and background pixels is unbalanced, as it requires sufficient labeled samples for model training. Supervised detection usually requires a lot of labeling work, which is costly and time-consuming in practical applications. However, when there are enough labeled samples and the target features are significant, it can provide high accuracy. Self-supervised detection methods use the structure of the data itself as a supervision signal without the need for external label information. In hyperspectral image processing, self-supervised learning can mine spectral changes in images to supplement prior knowledge, which is especially suitable for scenarios where prior spectral information is difficult to obtain. Unsupervised detection is a method that does not rely on labeled data. It identifies targets by analyzing the structure and patterns of the data itself, and is particularly suitable for situations where labeled data is difficult to obtain or too costly.
[0004] In recent years, with the vigorous development of deep learning technology, it has also been widely used in the field of hyperspectral target detection. Deep learning models have powerful feature learning capabilities and can automatically extract effective feature representations from a large amount of data, thereby improving the performance of target detection. However, the existing deep learning models applied to the field of hyperspectral target detection do not make sufficient use of computing resources, especially when processing large-scale data, which requires a large amount of computing power and memory usage. Summary of the Invention
[0005] To address the aforementioned problems and technical deficiencies, this application adopts the following technical solution: a hyperspectral target detection method based on depthwise separable convolution, comprising the following steps: In the low-level feature extraction stage, a depthwise separable convolution module is used to process the input hyperspectral image to obtain low-level features with preliminary feature information; Low-level features are input into a U-shaped Transformer for parameter optimization and adjustment to enhance feature representation. The enhanced features are then projected onto the joint coefficient space through the convolutional layer of the depthwise separable convolutional module, and then passed through a softmax function to achieve the constraint that the non-negative sum of the joint coefficients is equal to one. In the stage of decoding and restoring the input hyperspectral image, the background subspace is defined as a learnable variable, and a non-negativity constraint is obtained by activating a modified linear unit. The joint coefficients are multiplied by the joint subspace to complete the decoding and restoration of the input hyperspectral image.
[0006] Preferably, the depth-separable convolution module processes the input image by decomposing standard convolution into depthwise convolution and pointwise convolution; The depthwise separable convolution module breaks down the standard convolution operation in a convolutional layer into a depthwise convolution and a 1×1 pointwise convolution.
[0007] Furthermore, the depthwise convolution first passes through a convolution kernel with 3 kernels to generate 3 feature maps, and then passes through a convolution kernel with 4 kernels and a size of 1×1 to weight and combine the input feature maps in the depth direction to generate 4 feature maps. Depthwise convolution performs a convolution operation on each channel of the input feature map to extract spatial features without changing the number of channels. The specific process includes: Let the input feature map be ,in Indicates the height of the feature map. Indicates width, Indicates the number of channels; depthwise convolution uses... The size is convolution kernel ,in Each convolutional kernel corresponds to only one channel of the input feature map. The first in One channel Using convolution kernels Perform a convolution operation to obtain the output feature map. The first in One channel .
[0008] Furthermore, the specific calculation process of the depthwise convolution is as follows: Let the stride of the convolution be... Fill with For the output feature map The median coordinate is elements The calculation formula is as follows:
[0009] Among them, the input feature map Elements exceeding the boundary are processed according to the padding method, and a feature map is output. The size is ,in , .
[0010] Furthermore, the number of parameters in the depthwise convolution is Each channel corresponds to one The convolution kernel, depthwise convolution, reduces the number of parameters when extracting spatial features by reducing parameter sharing between different channels.
[0011] Furthermore, the pointwise convolution combines feature maps to generate new feature maps, extracts channel features from the input data, and uses... The convolution kernel performs a convolution operation on the output feature map of the depthwise convolution, linearly combining the channels of the feature map to change the number of channels; Let the output feature map of depthwise convolution be... Pointwise convolution is used The size is convolution kernel ,in This is the desired number of output channels for pointwise convolution output feature maps. The Middle One channel The coordinates are The formula for calculating the elements is:
[0012] That is, output feature map Each element in the diagram is a feature map output by a depthwise convolution. All channel elements at the corresponding position and the convolution kernel The corresponding channel The result of multiplying the elements and then summing them.
[0013] Furthermore, the number of parameters in the pointwise convolution is ,have indivual The convolutional kernels, each convolutional kernel has One parameter.
[0014] Furthermore, the total number of parameters in the depthwise separable convolution module is the sum of the number of depthwise convolution parameters and the number of pointwise convolution parameters, as shown in the following formula:
[0015] Standard convolution uses The size is The convolution kernel simultaneously processes the input feature map The convolution operation is performed on each channel, and the number of parameters is... .
[0016] Preferably, the parameter optimization adjustment is to improve the heads parameter of multi-head self-channel attention, so that the model can capture richer feature information by default.
[0017] A hyperspectral target detection system based on depthwise separable convolution includes: Subspace generation module for forming a non-negative joint subspace by connecting the target spectrum and the adaptively learnable background subspace; The encoding / decoding module is used for the subspace representation of the input hyperspectral image, which is represented by a deep subspace representation network and includes an encoder for extracting coefficients and a decoder for recovering the input. The synthesis and detection module is used to cover the synthesis of target components and the generation of detection maps.
[0018] Compared to existing technologies, the beneficial effects of this application are as follows: This application endows the model with depthwise separable convolution functionality through a depthwise separable convolution module, which further enhances the model's generalization ability. The depthwise separable convolution module can maintain high feature extraction capability while reducing the number of parameters, greatly improving the computational efficiency of the model. Especially when processing large-scale data, it can significantly reduce the amount of computation and memory usage. Attached Figure Description
[0019] In the attached diagram: Figure 1 This is a schematic diagram of the method steps in an embodiment of this application; Figure 2 This is an overall framework diagram of an embodiment of this application; Figure 3 This is a depthwise separable convolutional map according to an embodiment of this application; Figure 4 These are detection graphs of various comparison methods on the Sandiego dataset, based on embodiments of this application. Figure 5 These are detection graphs of various comparison methods on the abu-urban-1 dataset, representing embodiments of this application. Figure 6 These are detection graphs of various comparison methods on the abu-urban-2 dataset, based on embodiments of this application. Figure 7 This is a performance comparison table of embodiments of this application; Figure 8 The ROC curves of the DSC-HTD-IRN algorithm and four comparison algorithms in this application embodiment are shown on three datasets. Figure 9 This is a schematic diagram of the system structure according to an embodiment of this application. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments. Generally, the components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations.
[0021] Example 1 like Figure 1 As shown, the hyperspectral target detection method based on depthwise separable convolution includes the following steps: In the low-level feature extraction stage, a depthwise separable convolution module is used to process the input hyperspectral image to obtain low-level features with preliminary feature information; The depthwise separable convolution module processes the input image by decomposing the standard convolution into depthwise convolution and pointwise convolution; The depthwise separable convolution module breaks down the standard convolution operation in a convolutional layer into a depthwise convolution and a 1×1 pointwise convolution.
[0022] The depthwise convolution first passes through a convolution kernel with 3 kernels to generate 3 feature maps, and then passes through a convolution kernel with 4 kernels and a size of 1×1 to weight and combine the input feature maps in the depth direction to generate 4 feature maps. Depthwise convolution performs a convolution operation on each channel of the input feature map to extract spatial features without changing the number of channels. The specific process includes: Let the input feature map be ,in Indicates the height of the feature map. Indicates width, Indicates the number of channels; depthwise convolution uses... The size is convolution kernel ,in Each convolutional kernel corresponds to only one channel of the input feature map. The first in One channel Using convolution kernels Perform a convolution operation to obtain the output feature map. The first in One channel .
[0023] The specific calculation process of depthwise convolution is as follows: Let the stride of the convolution be... Fill with For the output feature map The median coordinate is elements The calculation formula is as follows:
[0024] Among them, the input feature map Elements exceeding the boundary are processed according to the padding method, and a feature map is output. The size is ,in , .
[0025] The number of parameters for depthwise convolution is Each channel corresponds to one The convolution kernel, depthwise convolution, reduces the number of parameters when extracting spatial features by reducing parameter sharing between different channels.
[0026] Pointwise convolution combines feature maps to generate new feature maps, extracting channel features from the input data. The convolution kernel performs a convolution operation on the output feature map of the depthwise convolution, linearly combining the channels of the feature map to change the number of channels; Let the output feature map of depthwise convolution be... Pointwise convolution is used The size is convolution kernel ,in This is the desired number of output channels for pointwise convolution output feature maps. The Middle One channel The coordinates are The formula for calculating the elements is:
[0027] That is, output feature map Each element in the diagram is a feature map output by a depthwise convolution. All channel elements at the corresponding position and the convolution kernel The corresponding channel The result of multiplying the elements and then summing them.
[0028] The number of parameters for pointwise convolution is ,have indivual The convolutional kernels, each convolutional kernel has One parameter.
[0029] The total number of parameters in a depthwise separable convolutional module is the sum of the number of parameters in the depthwise convolution and the number of parameters in the pointwise convolution, as shown in the following formula:
[0030] Standard convolution uses The size is The convolution kernel simultaneously processes the input feature map The convolution operation is performed on each channel, and the number of parameters is... .
[0031] Low-level features are input into a U-shaped Transformer for parameter optimization and adjustment to enhance feature representation. Parameter optimization tuning improves the head parameters of multi-head self-channel attention, enabling the model to capture richer feature information by default.
[0032] The enhanced features are projected onto the joint coefficient space through the convolutional layer of the depthwise separable convolutional module, and then passed through a softmax function to achieve the constraint that the non-negative sum of the joint coefficients is equal to one. In the stage of decoding and restoring the input hyperspectral image, the background subspace is defined as a learnable variable, and a non-negativity constraint is obtained by activating a modified linear unit. The joint coefficients are multiplied by the joint subspace to complete the decoding and restoration of the input hyperspectral image.
[0033] Example 2 The overall architecture consists of three critical paths: The first line represents the non-negative joint subspace formed by connecting the target spectrum and the adaptive, learnable background subspace; The second line is the subspace representation of the input hyperspectral image, which is the deep subspace representation network, containing an encoder for extracting coefficients and a decoder for recovering the input; The last line represents the detection block, which covers target component synthesis and detection map generation. The data flow of the improved overall architecture is described in detail below.
[0034] When the input hyperspectral image enters the system, in the low-level feature extraction stage, the traditional convolutional layer is no longer used. Instead, a depthwise separable convolutional module is used to process the input image.
[0035] The depthwise separable convolution module decomposes standard convolution into depthwise convolution and pointwise convolution. While reducing the number of parameters, it can maintain a high feature extraction capability and effectively improve the computational efficiency of the model. Especially when processing large-scale data, it can significantly reduce the amount of computation and memory usage. After processing by this module, low-level features with preliminary feature information are obtained.
[0036] Subsequently, these low-level features are fed into the U-shaped Transformer (Uformer). In the Uformer module, its parameters are optimized and adjusted, increasing the heads parameter of the multi-head self-channel attention (MHCA). This allows the model to capture richer feature information by default, enhancing the model's expressive power, better modeling long-distance feature interactions, and further enhancing feature representation.
[0037] Next, the enhanced features are projected onto the joint coefficient space through a convolutional layer that also uses a depthwise separable convolutional module, and then passed through a softmax function to achieve the constraint that the non-negative sum of the joint coefficients is equal to one.
[0038] During the decoding and recovery of the input hyperspectral image, the background subspace is still defined as a learnable variable and non-negativity is obtained through activation of a modified linear unit (ReLU).
[0039] The joint coefficients are multiplied by the joint subspace (the connection between the background subspace and the target spectrum) to complete the decoding and recovery of the input hyperspectral image.
[0040] Meanwhile, a depth parameter has been added to the initialization of the SRN class with a default value of 5. This makes the model architecture more flexible, and users can adjust the depth of the model according to specific needs to balance the complexity and performance of the model.
[0041] In operations parallel to the deep subspace representation network, the target coefficients are separated from the joint coefficients and then synthesized as target components by multiplying them by the target spectrum.
[0042] Finally, the target components are processed through a multi-scale convolutional layer to generate a detection map, such as... Figure 2 As shown.
[0043] In the task of hyperspectral target detection, a deep subspace representation network is used to integrate nonlinear feature representation with physical interpretability. The core objective of this network is to adaptively separate the target and background components in the hyperspectral image and demix them into subspaces (endmembers) and coefficients (abundance), thereby achieving accurate target detection.
[0044] Based on the traditional subspace representation model, it transforms the model into a deep learning network structure, which can effectively and conveniently use gradient descent to solve optimization problems.
[0045] To facilitate the separation of background and target, the original hyperspectral image can be decomposed into:
[0046] in Indicates background components, Indicates the target component, Indicates noise.
[0047] To accurately and effectively represent the background and target components, the formula can be reformulated by introducing subspace representation.
[0048] It is the background subspace. It is the background coefficient. It is the target coefficient. The prior known target spectrum is used as the target subspace. For the number of bands, Number of pixels yes The number of spectral signals contained therein, and satisfying .
[0049] Based on this, the detection problem can be transformed into the following optimization model:
[0050] The constraints are .
[0051] In the above formula, the first term This is the fidelity term, used to measure the difference between the decomposition result and the original hyperspectral image. The degree of difference; and These are the prior terms applied to the background and target coefficients, respectively. For example, sparse priors can make the coefficients sparse, low-rank priors can capture the linear correlation between coefficients, and smooth priors can ensure the local homogeneity of the background. and It is a trade-off parameter used to balance the effects of the fidelity term and the prior term; and They represent sizes of and A column vector of all 1s is used to ensure that the joint coefficients satisfy the constraint that the sum is 1.
[0052] Deep subspace representation networks mainly consist of an encoder and a decoder. The main task of the encoder is to learn the nonlinear mapping between the hyperspectral image and the joint coefficients, i.e.:
[0053] in This represents an encoder that estimates background and target coefficients from a hyperspectral image. This indicates the parameters involved in the encoder. This indicates a join operation.
[0054] Specifically, first apply a Convolution is used to extract low-level features, and a leaky ReLU activation is used to enhance non-linearity.
[0055] in It has The low-level characteristics of each channel, Indicates the number is of convolution, It is the leaky ReLU function.
[0056] Then, a lightweight multi-scale Transformer module, Uformer, is used for multi-scale analysis and feature enhancement:
[0057] in It is an enhanced feature.
[0058] Finally, the enhanced features are projected into the coefficient space by applying a convolution and a softmax function to obtain the desired joint coefficients:
[0059] It is important to note that the softmax function is applied along the channel dimension to ensure that the non-negativity of the joint coefficients is a constraint. The reshape operation here expands the tensor into a matrix form.
[0060] The decoder recovers the hyperspectral image from the joint coefficients by multiplying by the joint subspace, which can be expressed as:
[0061] This is equivalent to:
[0062] in It is an adaptive learnable variable, the ReLU function. To ensure the non-negativity of the background subspace, the reshape operation here folds the matrix into tensor form, resulting in the restored hyperspectral image. .
[0063] In the field of hyperspectral target detection, although traditional convolution operations can effectively extract features, they have a large number of parameters and high computational cost, posing challenges when processing large-scale hyperspectral data.
[0064] To address this issue, depthwise separable convolution is introduced into the improved model. Depthwise separable convolution decomposes the standard convolution operation in some convolutional layers into one depthwise convolution and one 1×1 pointwise convolution. This approach significantly reduces computation and the number of parameters at the cost of slightly sacrificing accuracy, thereby improving speed and reducing the memory required for weight parameter files. A schematic diagram of the depthwise separable convolution used for a three-channel input image is shown below. Figure 3 As shown, for the same 3-channel input image, depthwise convolution first passes through a convolution kernel with 3 filters, generating 3 feature maps. Then, it passes through a convolution kernel with 4 kernels and a size of 1×1, which weights and combines the feature maps from the previous step in the depth direction to generate 4 feature maps, with the same output dimension as regular convolution.
[0065] Depthwise convolution is the first step in depthwise separable convolution. Its core idea is to perform convolution operations on each channel of the input feature map separately, thereby extracting spatial features without changing the number of channels.
[0066] A single kernel in a depthwise convolution is responsible for one channel, used to extract spatial features from the data. The number of feature maps generated is the same as the number of channels in the input data. The output is the feature map corresponding to each channel. Depthwise convolution only changes the size of the feature map without changing the number of channels in the feature map.
[0067] Let the input feature map be ,in Indicates the height of the feature map. Indicates width, Indicates the number of channels.
[0068] Depth convolution The size is convolution kernel ,in Each convolutional kernel corresponds to only one channel of the input feature map.
[0069] For the input feature map The first in One channel Using convolution kernels Perform a convolution operation to obtain the output feature map. The first in One channel The specific calculation process is as follows: Let the stride of the convolution be... Fill with For the output feature map The median coordinate is elements The calculation formula is as follows:
[0070] Among them, the input feature map Elements exceeding the boundary are processed according to the padding method (e.g., zero padding). Output feature map. The size is ,in , .
[0071] The number of parameters for depthwise convolution is Because each channel corresponds to one Compared to standard convolution, depthwise convolution reduces parameter sharing between different channels when extracting spatial features, thus reducing the number of parameters.
[0072] Since depthwise convolution performs convolution operations on each channel of the input data independently, it does not effectively utilize the information of different channels at the same spatial location. Therefore, pointwise convolution is needed to combine these feature maps and generate new feature maps to extract the channel features of the input data.
[0073] Pointwise convolution is the second step in depthwise separable convolution; it uses... The convolution kernel performs a convolution operation on the output feature map of the depthwise convolution. The main purpose is to linearly combine the channels of the feature map, thereby changing the number of channels.
[0074] Let the output feature map of depthwise convolution be... Pointwise convolution is used The size is convolution kernel ,in This is the desired number of output channels.
[0075] For point-by-point convolution output feature map The Middle One channel The coordinates are The formula for calculating the elements is:
[0076] That is, output feature map Each element in the diagram is a feature map output by a depthwise convolution. All channel elements at the corresponding position and the convolution kernel The corresponding channel The result of multiplying the elements and then summing them.
[0077] The number of parameters for pointwise convolution is Because there is indivual The convolutional kernels, each convolutional kernel has One parameter.
[0078] The total number of parameters in a depthwise separable convolution is the sum of the number of parameters in a depthwise convolution and the number of parameters in a pointwise convolution.
[0079] In contrast, standard convolution uses The size is The convolution kernel simultaneously processes the input feature map The convolution operation is performed on each channel, and the number of parameters is... .
[0080] when Larger and In this case, depthwise separable convolution has far fewer parameters than standard convolution, which can significantly reduce computation and memory usage.
[0081] For example, when , , At that time, the number of parameters in standard convolution is The number of parameters for depthwise separable convolution is It is only about 12% of the number of parameters in a standard convolution.
[0082] Example 3 Experiments were conducted on three benchmark datasets: San Diego, HYDICE, and Urban, to verify the effectiveness of the proposed method.
[0083] Figure 5 (a) and (e) show the false-color image and the real map of the dataset, respectively.
[0084] Since the target spectrum should be known a priori, in our experiment, we selected the pixel with the smallest average distance from other target pixels as the target spectrum.
[0085] The San Diego dataset was acquired by the Airborne Visible / Infrared Imaging Spectroradiometer (AVIRIS) over San Diego International Airport in California, USA. The sub-scenes used consisted of 100×100 pixels, with a spatial resolution of 3.5 meters. After removing low signal-to-noise ratio (SNR) and water vapor absorption bands (1–6, 33–35, 97, 107–113, 153–166, and 221–224), 189 bands remained in the 400–2500 nm range, with a spectral resolution of 10 nm. Three aircraft were the targets to be detected, comprising 58 pixels in the sub-scene dataset.
[0086] Texas Coast Datasets: The abu-urban-1 and abu-urban-2 datasets are located on the coast of Texas, USA. They have a spatial size and resolution of 100×100 pixels and 17.2 m. After removing low signal-to-noise ratio bands, the abu-urban-1 and abu-urban-2 datasets retain 207 and 204 bands, respectively. Outliers are buildings in the data.
[0087] To evaluate the performance of the detection algorithm, four commonly used algorithms were compared: LRASR, GTVLRR, PTA, and HTD-IRN. Since this method is unsupervised and designed for object detection based on the entire image, it does not require the construction of training and test sets. For each dataset, the entire image is used as input, with a batch size of 1. The learning rate is initially set to 0.001 and decays by half every 1000 iterations.
[0088] All experiments were conducted using Python 3.10.14, PyTorch 2.0.1, and CUDA 11.7, with an Intel(R) Core(TM) i7-12700H CPU and an NVIDIA GeForce RTX 4050 Laptop GPU server.
[0089] The results were obtained by comparing the visual detection map, ROC curve, and AUC value. The visual detection map is an important means to intuitively show the detection results of the model. By comparing the detection map output by the model with the real target distribution, the detection status of the model in each region can be directly observed.
[0090] ROC curves are a common tool for evaluating the performance of classification models. They depict the detection probability, ) and false alarm rate The relationship between the two values demonstrates the model's classification ability. The detection probability, also known as the true positive rate, represents the proportion of target pixels correctly detected by the model to the actual number of target pixels. The false alarm rate represents the proportion of background pixels that the model incorrectly detects as targets to the actual number of background pixels.
[0091] Both can be calculated using the following formula:
[0092] in, (True Positives) represents the number of true positive samples, that is, the number of pixels that the model correctly detects as targets; (False Negatives) represents the number of false negative samples, which are the number of pixels that were actually targets but were not detected by the model. (False Positives) represents the number of false positive samples, that is, the number of background pixels that the model mistakenly detected as targets; (TrueNegatives) represents the number of true negative samples, which is the number of pixels that the model correctly detects as background.
[0093] false alarm rate The horizontal axis represents the detection probability. Using the vertical axis as the ordinate, connecting these points forms the ROC curve.
[0094] The closer the ROC curve is to the top left corner, the better the model's performance, meaning it can achieve a high detection probability with a low false alarm rate. The AUC value is the area under the ROC curve. It is a comprehensive evaluation metric used to measure the overall classification performance of the model. The AUC value ranges from [0, 1], and the closer the value is to 1, the better the model's performance.
[0095] In the method proposed in this application, m is the number of background endmembers, used to control the complexity of the background subspace, and is set to 30; the regularization parameter of the sparse loss... The initial learning rate is used to control the strength of the sparse constraint and is set to 10; the initial learning rate is used for the optimizer's parameter updates and is set to 0.001.
[0096] Figure 4 In the image, (a) is a pseudo-color image, (b) is ground truth, (c) is LRASR, (d) is GTVLRR, (e) is PTA, (f) is HTD-IRN, and (g) is DSC-HTD-IRN.
[0097] Figure 5In the image, (a) is a pseudo-color image, (b) is ground truth, (c) is LRASR, (d) is GTVLRR, (e) is PTA, (f) is HTD-IRN, and (g) is DSC-HTD-IRN.
[0098] Figure 6 In the image, (a) is a pseudo-color image, (b) is ground truth, (c) is LRASR, (d) is GTVLRR, (e) is PTA, (f) is HTD-IRN, and (g) is DSC-HTD-IRN.
[0099] Figure 4-6 The detection maps obtained by five contrast methods on three benchmark datasets are shown.
[0100] It can be observed that, compared with its competitors, DSC-HTD-IRN achieves a satisfactory balance in highlighting the target and suppressing the background.
[0101] For example, in the Sandiego dataset, it is difficult to detect the top aircraft because some dim and small targets are located within it.
[0102] from Figure 4 As can be seen from (c), (d), and (e), LRASR, GTVLRR, and PTA all exhibited false alarms to some extent. The HTD-IRN detection results showed that the target outline was incomplete and unclear. DSC-HTD-IRN, while suppressing background pixels, effectively displayed a clear and complete target outline.
[0103] For example, in the abu-urban-2 urban dataset, target pixels account for a large proportion of the entire scene, from Figure 6 As can be seen from (c), (d), and (f), LRASR, GTVLRR, and HTD-IRN lose many target pixels. Figure 6 (d) and (e) show that GTVLRR and PTA have some shortcomings in suppressing background pixels, and the target edges detected by these comparison methods are blurry. The third method, DSC-HTD-IRN, shows relatively good performance in both suppressing background pixels and highlighting target pixels, and the detected targets are clear and distinct.
[0104] Figure 7 shows the ROC curves of the five methods on three datasets. It can be seen that the ROC curve of DSC-HTD-IRN is almost in the upper left corner on all three datasets, which indicates that it has good and stable performance.
[0105] Figure 7 for This table shows a performance comparison of DSC-HTD-IRN and four other methods on three datasets. The best results are indicated in bold, and the second-best results are indicated by underline. It can be seen that the DSC-HTD-IRN method performs better on each dataset. All are the highest.
[0106] Overall, DSC-HTD-IRN is more reliable than the other four comparison methods.
[0107] Figure 7 In Performance comparison of DSC-HTD-IRN and four contrasting methods on three datasets. Bold data indicates the best result, and underlined data indicates the second best result. Figure 8 The first row shows the overall view, and the second row shows a partial view to observe the specific differences.
[0108] By utilizing depthwise separable convolutions, the number of computational parameters is reduced, improving detection performance and model generalization ability. Experiments conducted on three benchmark datasets and compared with four other algorithms validate the effectiveness and stability of the proposed method. The final experimental results, including visualized detection maps, ROC curves, and AUC values, demonstrate that the proposed method is a promising solution for hyperspectral object detection and outperforms other competing methods.
[0109] Example 4 like Figure 9 As shown, a hyperspectral target detection system based on depthwise separable convolution includes: Subspace generation module for forming a non-negative joint subspace by connecting the target spectrum and the adaptively learnable background subspace; The encoding / decoding module is used for the subspace representation of the input hyperspectral image, which is represented by a deep subspace representation network and includes an encoder for extracting coefficients and a decoder for recovering the input. The synthesis and detection module is used to cover the synthesis of target components and the generation of detection maps.
[0110] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications, improvements, and substitutions without departing from the concept of this application, and these all fall within the protection scope of this application.
Claims
1. A hyperspectral target detection method based on depthwise separable convolution, characterized in that, Includes the following steps: In the low-level feature extraction stage, a depthwise separable convolution module is used to process the input hyperspectral image to obtain low-level features with preliminary feature information; Low-level features are input into a U-shaped Transformer for parameter optimization and adjustment to enhance feature representation. The enhanced features are then projected onto the joint coefficient space through the convolutional layer of the depthwise separable convolutional module, and then passed through a softmax function to achieve the constraint that the non-negative sum of the joint coefficients is equal to one. In the stage of decoding and restoring the input hyperspectral image, the background subspace is defined as a learnable variable, and a non-negativity constraint is obtained by activating a modified linear unit. The joint coefficients are multiplied by the joint subspace to complete the decoding and restoration of the input hyperspectral image.
2. The hyperspectral target detection method based on depthwise separable convolution according to claim 1, characterized in that, The depth-separable convolution module processes the input image by decomposing standard convolution into depthwise convolution and pointwise convolution; The depthwise separable convolution module breaks down the standard convolution operation in a convolutional layer into a depthwise convolution and a 1×1 pointwise convolution.
3. The hyperspectral target detection method based on depthwise separable convolution according to claim 2, characterized in that, The depthwise convolution first passes through a convolution kernel with 3 kernels to generate 3 feature maps, and then passes through a convolution kernel with 4 kernels and a size of 1×1 to weight and combine the input feature maps in the depth direction to generate 4 feature maps. Depthwise convolution performs a convolution operation on each channel of the input feature map to extract spatial features without changing the number of channels. The specific process includes: Let the input feature map be ,in Indicates the height of the feature map. Indicates width, Indicates the number of channels; depthwise convolution uses... The size is convolution kernel ,in Each convolutional kernel corresponds to only one channel of the input feature map. The first in One channel Using convolution kernels Perform a convolution operation to obtain the output feature map. The first in One channel .
4. The hyperspectral target detection method based on depthwise separable convolution according to claim 3, characterized in that, The specific calculation process for the depthwise convolution is as follows: Let the stride of the convolution be... Fill with For the output feature map The median coordinate is elements The calculation formula is as follows: Among them, the input feature map Elements exceeding the boundary are processed according to the padding method, and a feature map is output. The size is ,in , .
5. The hyperspectral target detection method based on depthwise separable convolution according to claim 4, characterized in that, The number of parameters for the depthwise convolution is Each channel corresponds to one The convolution kernel, depthwise convolution, reduces the number of parameters when extracting spatial features by reducing parameter sharing between different channels.
6. The hyperspectral target detection method based on depthwise separable convolution according to claim 2, characterized in that, The pointwise convolution combines feature maps to generate new feature maps, extracting channel features from the input data, and using... The convolution kernel performs a convolution operation on the output feature map of the depthwise convolution, linearly combining the channels of the feature map to change the number of channels; Let the output feature map of depthwise convolution be... Pointwise convolution is used The size is convolution kernel ,in This is the desired number of output channels for pointwise convolution output feature maps. The Middle One channel The coordinates are The formula for calculating the elements is: That is, output feature map Each element in the diagram is a feature map output by a depthwise convolution. All channel elements at the corresponding position and the convolution kernel The corresponding channel The result of multiplying the elements and then summing them.
7. The hyperspectral target detection method based on depthwise separable convolution according to claim 6, characterized in that, The number of parameters for the pointwise convolution is ,have indivual The convolutional kernels, each convolutional kernel has One parameter.
8. The hyperspectral target detection method based on depthwise separable convolution according to claim 2, characterized in that, The total number of parameters in the depthwise separable convolution module is the sum of the number of depthwise convolution parameters and the number of pointwise convolution parameters, as shown in the following formula: Standard convolution uses The size is The convolution kernel simultaneously processes the input feature map The convolution operation is performed on each channel, and the number of parameters is... .
9. The hyperspectral target detection method based on depthwise separable convolution according to claim 1, characterized in that, The parameter optimization adjustment refers to improving the heads parameter of the multi-head self-channel attention, enabling the model to capture richer feature information by default.
10. A hyperspectral target detection system based on depthwise separable convolution, characterized in that, include: Subspace generation module for forming a non-negative joint subspace by connecting the target spectrum and the adaptively learnable background subspace; The encoding / decoding module is used for the subspace representation of the input hyperspectral image, which is represented by a deep subspace representation network and includes an encoder for extracting coefficients and a decoder for recovering the input. The synthesis and detection module is used to cover the synthesis of target components and the generation of detection maps.