Methods, devices, media, and equipment for identifying camouflaged targets based on UAV imagery

Abstract

This application discloses a method, apparatus, medium, and device for camouflage target recognition based on UAV imagery, relating to the field of image recognition technology. The method includes: performing image recognition on the UAV imagery to be camouflaged to obtain a target image in the near-infrared band of the target; extracting features from the target image using a preset spatial spectral autoencoder to obtain spatial spectral features corresponding to the target image; and classifying the pixels of the target image using a preset graph convolutional network classifier based on the spatial spectral features to obtain the camouflage target recognition result of the UAV imagery. The method of this application can improve the accuracy of camouflage target recognition from UAV imagery.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method, apparatus, medium, and device for identifying camouflaged targets based on UAV imagery. Background Technology

[0002] Accurately locating and detecting camouflaged targets against complex natural vegetation backgrounds is a key challenge in intelligent monitoring and environmental security applications. Typically, camouflage mimics the spectral reflectance characteristics and texture of the background within specific wavelength ranges (such as visible light and mid-infrared), effectively concealing the target from visual and panchromatic sensors. However, with the continuous development of hyperspectral imaging technology, the adversarial landscape of target detection and material camouflage is constantly being reshaped. Hyperspectral imaging technology can acquire spatial and spectral information encompassing hundreds of continuous narrow bands, demonstrating enormous application potential in many fields such as environmental monitoring, resource exploration, and agricultural assessment. This technology provides crucial clues for camouflage net target detection by identifying subtle spectral differences between camouflage materials and natural vegetation. Nevertheless, due to the high dimensionality and nonlinearity of hyperspectral data, as well as the significant spectral similarity between targets and backgrounds, accurately detecting camouflaged targets in hyperspectral imagery remains a significant challenge.

[0003] Traditional hyperspectral target detection methods rely on manually designed features, such as GIST, LBP, and HOG. These methods have poor versatility and limited detection performance. Machine learning methods, such as support vector machines and random forests, have improved the ability to process high-dimensional data, but their performance is still limited by the quality of manually designed features and often ignores spatial context information. Deep neural networks, such as convolutional neural networks (CNNs), autoencoders (SEs), and recurrent neural networks (RNNs), can extract hierarchical spatial-spectral feature information from hyperspectral images through end-to-end learning, overcoming the shortcomings of manually designed features. Subsequently, architectures such as 3D-CNN, Transformer, and Mamba have further promoted spatial-spectral fusion and long-sequence spectral modeling. However, these methods are usually based on feature extraction and modeling across the entire spectral band, ignoring the spectral confusion between the target and the background in specific bands. In particular, camouflaged targets are designed to simulate the spectral characteristics of their surrounding environment, exhibiting high spectral similarity in specific bands. Without an effective discriminative band selection mechanism, the model will overfit on irrelevant or specific bands, thereby reducing its generalization ability. Furthermore, these supervised models rely on large amounts of pixel-level labeled data, a requirement that poses a significant bottleneck in practical applications. Although semi-supervised and unsupervised methods have alleviated labeling costs to some extent in recent years, they are often inconsistent with the ultimate detection task objectives and have also failed to effectively address the problem of band redundancy. Summary of the Invention

[0004] In view of this, the present invention provides a method, apparatus, medium and device for identifying camouflaged targets based on UAV imagery. The main purpose is to solve the problem of inaccurate camouflage target identification caused by the limited ability to select and focus discriminative bands and the failure to fully utilize the spatial-spectral joint features of pixels.

[0005] To address the aforementioned problems, this application provides a method for identifying camouflaged targets based on UAV imagery, comprising: Image recognition is performed on the image of the drone image to be disguised to obtain a target image in the near-infrared band of the drone image to be disguised to be identified. The target image is feature extracted using a preset spatial spectral autoencoder to obtain the spatial spectral features corresponding to the target image; Based on the spatial spectral features, a preset graph convolutional network classifier is used to classify the pixels of the target image to obtain the camouflage target recognition result of the UAV image to be camouflaged.

[0006] Optionally, the step of performing image recognition on the image of the UAV to be camouflaged to obtain a target image in the near-infrared band of the image of the UAV to be camouflaged specifically includes: The image of the drone to be disguised is identified to obtain the target area and the background area; The average reflectance of the target area and the background area in different bands are calculated respectively to obtain the first average reflectance of the target area in different bands and the second average reflectance of the background area in different bands. The spectral difference value of the same band is obtained by calculating the first average reflectance and the second average reflectance based on the same band. The near-infrared band of the target to be identified in the UAV image is determined based on the spectral difference values ​​of different bands. The target image is obtained by extracting the image corresponding to the near-infrared band of the target in the image of the UAV that is to be disguised.

[0007] Optionally, the step of using a preset spatial spectral autoencoder to extract features from the target image to obtain spatial spectral features corresponding to the target image specifically includes: Spatial spectral encoding is performed on the target image to obtain spectral groups corresponding to each pixel of the target image; The joint spatial-spectral features of local regions of each of the aforementioned spectral groups are encoded to obtain global spatial-spectral encoded features; The global spatial spectrum coding features are decoded to obtain the spatial spectrum features corresponding to the target image.

[0008] Optionally, encoding the joint spatial-spectral features of local regions of each of the spectral groups to obtain global spatial-spectral encoded features specifically includes: The preset spatial spectral autoencoder is used to perform vector transformation on the spectral groups to obtain a first feature vector corresponding to each spectral group. Position encoding is performed on each of the first feature vectors to obtain a second feature vector corresponding to each of the spectral groups; The second feature vector is extracted using the Transformer encoding layer of the preset spatial spectral autoencoder to obtain a third feature vector corresponding to each of the spectral groups. The third feature vectors are subjected to feature fusion and pooling to obtain the global spatial spectrum coding features; The location coding includes spatial location coding and spectral group location coding.

[0009] Optionally, the step of using the Transformer encoding layer of the preset spatial spectral autoencoder to extract features from the second feature vector to obtain a third feature vector corresponding to each of the spectral groups specifically includes: A multi-head self-attention network with a Transformer coding layer is used to extract features from the second feature vector to obtain the fourth feature vector; Based on the second feature vector and the fourth feature vector, residual linking and layer normalization are performed to obtain the fifth feature vector; The feedforward network of the Transformer coding layer is used to extract features from the fifth feature vector to obtain the sixth feature vector; Based on the fourth and sixth feature vectors, residual linking and layer normalization are performed to obtain the third feature vector corresponding to each of the spectral groups.

[0010] Optionally, the step of classifying the pixels of the target image using a preset graph convolutional network classifier based on the spatial spectral features to obtain the camouflage target recognition result of the UAV image to be camouflaged specifically includes: The graph data structure of the target image is constructed by using the pixels of the target image as graph nodes, the spatial spectral features as node features of the pixels, and the connection lines of the neighboring pixels corresponding to the pixels obtained by cosine similarity calculation as edges. Based on the graph data structure, a neighborhood aggregation method is used to perform graph convolution calculation to obtain target node feature vectors with enhanced spatial context, connectivity, and boundary integrity. The preset graph convolutional network classifier is used to classify the feature vector of the target node to obtain the pixel probability value of the target image, so as to obtain the camouflage target recognition result of the drone image to be camouflaged.

[0011] Optionally, the step of constructing the graph data structure of the target image using pixels of the target image as graph nodes, using the spatial spectral features as node features of the pixels, and using the connection lines between the neighboring pixels corresponding to the pixel obtained by cosine similarity calculation as edges, specifically includes: The pixels of the target image are defined as graph nodes of the graph data structure to be constructed; The spatial spectral feature is defined as the node feature of the graph node; Calculate the cosine similarity between each of the aforementioned pixels; The neighboring nodes of the pixel are filtered based on the cosine similarity of each pixel to obtain the sparse adjacency matrix corresponding to the pixel. The sparse adjacency matrix is ​​symmetrically normalized to construct the graph data structure of the target image.

[0012] To address the aforementioned technical problems, this application provides a camouflaged target identification device based on UAV imagery, comprising: The identification module is used to identify the image of the drone to be disguised and obtain a target image in the near-infrared band of the image of the drone to be disguised. The feature extraction module is used to extract features from the target image using a preset spatial spectral autoencoder to obtain spatial spectral features corresponding to the target image; The classification module is used to classify the pixels of the target image based on the spatial spectral features using a preset graph convolutional network classifier, so as to obtain the camouflage target recognition result of the UAV image to be camouflaged.

[0013] To address the aforementioned technical problems, this application provides a storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned method for identifying camouflaged targets based on UAV imagery.

[0014] To address the aforementioned technical problems, this application provides an electronic device, comprising at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the aforementioned method for identifying camouflaged targets based on UAV imagery.

[0015] The beneficial effects of this application are as follows: Based on UAV hyperspectral imagery, this application improves model recognition efficiency and classification accuracy by deeply mining spectral differences between bands and focusing on a subset of bands with strong discriminative power. Building upon this, a spatial-spectral coding module is designed and combined with the Transformer mechanism to effectively establish contextual dependencies, achieving deep modeling of spatial-spectral features. By constructing a graph structure and introducing a graph convolutional network, accurate and efficient detection of camouflaged targets is achieved. This application improves the accuracy of camouflaged target recognition under conditions of high similarity between the target and the background.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a method for identifying camouflaged targets based on UAV imagery provided in an embodiment of this application is shown. Figure 2 A flowchart illustrating a method for identifying camouflaged targets based on UAV imagery, according to another embodiment of this application, is shown. Figure 3 This illustration shows a structural block diagram of a camouflage target identification device based on UAV imagery, according to another embodiment of this application. Figure 4 The diagram illustrates the comparison results of various comparison methods according to embodiments of this application on a self-built camouflaged target dataset. Figure 4 (a) A schematic diagram of the original input image of an embodiment of this application is shown; Figure 4 (b) A schematic diagram of a real label according to an embodiment of this application is shown; Figure 4 (c) A schematic diagram showing the recognition results obtained by the Minimum Distance Classifier (MDC) method according to an embodiment of this application; Figure 4 (d) shows a schematic diagram of the recognition results obtained using the Support Vector Machine (SVM) method in an embodiment of this application; Figure 4 (e) shows a schematic diagram of the recognition result obtained by using the Convolutional Neural Network (CNN) method according to an embodiment of this application; Figure 4 (f) shows a schematic diagram of the recognition results obtained by the HyBridSN method of the hybrid spectral spatial network according to an embodiment of this application; Figure 4 (g) shows a schematic diagram of the recognition results obtained by using the camouflaged target recognition method based on UAV imagery of this application. Detailed Implementation

[0018] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0019] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0020] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0021] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0022] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0023] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0024] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0025] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0026] This application provides a method for identifying camouflaged targets based on UAV imagery, such as... Figure 1 As shown, it includes: Step S101: Perform image recognition on the image of the UAV to be disguised target recognition to obtain the target image in the near-infrared band of the UAV image to be disguised target recognition; In this step, the image of the drone to be camouflaged is identified to obtain the target area and the background area; the average reflectance of the target area and the background area in different bands is calculated to obtain the first average reflectance of the target area and the second average reflectance of the background area in different bands; the first average reflectance and the second average reflectance of the same band are processed to obtain the spectral difference value of the same band; the near-infrared band of the target to be camouflaged is determined based on the spectral difference value of different bands; the image corresponding to the near-infrared band of the target to be camouflaged is extracted from the image of the drone to be camouflaged to obtain the target image.

[0027] Step S102: Use a preset spatial spectral autoencoder to extract features from the target image to obtain the spatial spectral features corresponding to the target image; In this step, spatial spectral encoding is performed on the target image to obtain spectral groups corresponding to each pixel of the target image; the spatial-spectral joint features of local regions of each spectral group are encoded to obtain global spatial-spectral encoded features; and the global spatial-spectral encoded features are decoded to obtain the spatial-spectral features corresponding to the target image.

[0028] Step S103: Based on the spatial spectral features, a preset graph convolutional network classifier is used to classify the pixels of the target image to obtain the camouflage target recognition result of the UAV image to be camouflaged.

[0029] In this step, the target image's pixels are used as graph nodes, the spatial spectral features are used as node features of the pixels, and the connection lines between the corresponding neighboring pixels obtained by cosine similarity calculation are used as edges to construct the target image's graph data structure. Based on the graph data structure, a neighborhood aggregation method is used to perform graph convolution calculation to obtain target node feature vectors with enhanced spatial context, connectivity, and boundary integrity. The preset graph convolutional network classifier is used to classify the target node feature vectors to obtain the pixel probability values ​​of the target image, thereby obtaining the camouflage target recognition result of the UAV image to be camouflaged.

[0030] This application leverages UAV hyperspectral imagery to deeply mine spectral differences between bands, focusing on a subset of bands with strong discriminative power, thereby improving model recognition efficiency and classification accuracy. Building upon this, a spatial-spectral coding module is designed and combined with the Transformer mechanism to effectively establish contextual dependencies, achieving deep modeling of spatial-spectral features. By constructing a graph structure and introducing a graph convolutional network, accurate and efficient detection of camouflaged targets is achieved. This application improves the accuracy of camouflaged target recognition under conditions of high similarity between the target and the background.

[0031] Another embodiment of this application provides a different method for identifying camouflaged targets based on UAV imagery, such as... Figure 2 As shown, it includes: Step S201: Perform image recognition on the image of the UAV to be disguised target to obtain a target image in the near-infrared band of the image of the UAV to be disguised target; In this step, the image of the UAV used to identify the target to be camouflaged is recognized to obtain the target area and the background area. The average reflectance of the target area and the background area in different bands is calculated to obtain the first average reflectance of the target area and the second average reflectance of the background area in different bands. Based on the first average reflectance and the second average reflectance of the same band, the spectral difference value of the same band is obtained. The mathematical formula for the calculation is as follows:

[0032] in, The first average reflectivity of the target area under the λ band; This represents the second average reflectance of the background region under the λ band. The spectral difference value is defined as the λ band. Based on the spectral difference values ​​of different bands, the target near-infrared band for identification of the UAV image to be disguised is determined; the image corresponding to the target near-infrared band in the UAV image to be disguised is extracted to obtain the target image. A higher value indicates stronger discrimination between the target and background in that band; conversely, a lower value indicates higher similarity. Analysis shows that in the visible light band of 450-650 nm, the spectral difference value is generally below 0.05, confirming the effectiveness of camouflage simulation in this region. The most discriminative band is concentrated in the near-infrared region after 708 nm, with a difference value as high as 0.1856. This makes visible light-based detection methods prone to failure, while utilizing non-visible light bands such as near-infrared allows for efficient and reliable detection of camouflaged targets. Therefore, we selected the 708 nm-900 nm near-infrared band as the discriminative band for model input.

[0033] Step S202: Perform spatial spectral encoding on the target image to obtain spectral groups corresponding to each pixel of the target image; In this step, the neighboring pixels of each pixel in the target image are extracted to obtain the corresponding neighborhood spatial pixel block. A sliding window of size K×K can be used for pixel sampling to obtain the neighborhood spatial pixel block; the size of K can be set according to the pixel size of the actual image. The neighborhood spatial pixel block carries contextual information and spectral information within its neighborhood. The neighborhood spatial pixel block is divided into a predetermined number of non-overlapping spectral groups along the spectral dimension to obtain the spectral group corresponding to each pixel. Specifically, to fully capture the contextual information of each pixel in the hyperspectral image, a local neighborhood is first constructed for each pixel to mine the spatial correlation between adjacent pixels. For hyperspectral images... H is the image height; W is the image width; C is the number of image channels; and represents each pixel in the image. Extract its K×K neighborhood pixel blocks. The mathematical expression is as follows:

[0034] Where r is the neighborhood radius, , k It is an odd number. Each neighborhood spatial pixel block It simultaneously includes both spatial context information and spectral information within the neighborhood. Furthermore, to deeply mine the spectral features of each band of the hyperspectral pixel and long-range contextual associations, each neighborhood spatial pixel block... Dividing the spectral space into G non-overlapping spectral groups along the spectral dimension allows the model to better understand the local correlations in different spectral intervals: The mathematical expression for representing neighborhood pixel blocks using spectral groups is as follows:

[0035] Among them, the Each spectral group is , This represents the number of bands included in each group.

[0036] Step S203: Encode the joint spatial-spectral features of the local regions of each of the spectral groups to obtain global spatial-spectral encoded features; In the specific implementation process of this step, the group embedding layer of the preset spatial spectral autoencoder is used to perform vector transformation on the spectral groups to obtain the first feature vector corresponding to each spectral group; for each spectral group... First, the spectral group is flattened in space to form a space-spectral joint vector, the mathematical expression of which is as follows:

[0037] in, It is a joint spatial-spectral vector; Let be the number of spatial locations. A linear projection method is used to perform feature space mapping on the spatial-spectral joint vector to obtain the first feature vector corresponding to each spectral group; the mathematical expression is as follows:

[0038] in, and For the first The learnable parameters of the group; This represents the model dimension, specifically the hidden layer dimension in the Transformer, and can be set to 128. For the first The first feature vector of the spectral group. Position encoding is performed on each of the first feature vectors to obtain a second feature vector corresponding to each spectral group; spatial position embedding is performed on the first feature vectors to obtain a spatial position encoded vector, the mathematical expression of which is as follows:

[0039] in, It is a spatial location encoding vector.

[0040] Perform spectral group position embedding on the first feature vector to obtain the spectral group position encoding vector; the mathematical expression is as follows:

[0041] in: This is the location encoding vector for the spectral group.

[0042] S represents the number of spatial locations, and G represents the number of spectral groups, each spectral group Each spatial location (x, y) is assigned a learnable vector to distinguish between different groups. Each spatial location is mapped to a unique learnable vector via an index.

[0043] The second feature vector is obtained by performing an addition operation on the spatial location encoding vector, the spectral group location encoding vector, and the first feature vector. The mathematical expression is as follows:

[0044] in, This represents the second feature vector after spatial location encoding and spectral group location encoding of the first feature vector.

[0045] The second feature vector is extracted using the Transformer encoding layer of the preset spatial spectral autoencoder to obtain a third feature vector corresponding to each of the spectral groups. A multi-head self-attention network with Transformer coding layers is used to extract features from the second feature vector to obtain the fourth feature vector; the mathematical expression is as follows:

[0046] in, This is the second feature vector; Represents the fourth eigenvector; The output projection matrix is ​​used to map the results of the multi-head attention mechanism back to the original model dimension. for h A parallel attention head; Each attention point is calculated as follows:

[0047]

[0048] Where: Q is the query matrix, used to match information from other locations; K is the key matrix, used for matching the query; V is the value matrix, used to store the feature information to be extracted; T is the transpose operation; The dimension of the key vector is used to scale the scores to avoid excessively large values. For the first i indivual Q Weight; For the first i indivual K Weight; For the first i indivual V Weights; based on the second and fourth feature vectors, residual linking and layer normalization are performed to obtain the fifth feature vector; the feedforward network of the Transformer coding layer is used to extract features from the fifth feature vector to obtain the sixth feature vector; the mathematical expression is as follows:

[0049] in, This is the sixth eigenvector; These are the weights of the first linear layer of the feedforward network. This is the bias of the first linear layer of the feedforward network; These are the weights of the second linear layer in the feedforward network. This is the bias of the second linear layer of the feedforward network; It is a non-linear activation function.

[0050] Based on the fourth and sixth eigenvectors, residual linking and layer normalization are performed to obtain the third eigenvector corresponding to each spectral group. The mathematical expression is as follows:

[0051] in, The intermediate features output from the attention module are processed by residual connections and layer normalization. For the first l The layer is the feature vector output by the Transformer encoding layer; Layer Norm is the layer normalization function.

[0052] The third feature vectors are fused and pooled to obtain the global spatial spectrum coding features; after feature extraction by an L-layer Transformer encoder, we obtain the spectral and spatial location features of all groups. To obtain a global representation of the samples, the mathematical expression for global average pooling is as follows:

[0053] The location coding includes spatial location coding and spectral group location coding; The characteristic mean of all spectral groups and spatial locations; The i-th feature vector output by the L-th Transformer layer contains all spectral groups and spatial locations; N is the number of spatial locations within a single spectral group; G is the number of spectral groups, which divides the entire hyperspectral image into spectral groups. i For location index.

[0054] Step S204: Decode the global spatial spectrum coding features to obtain the spatial spectrum features corresponding to the target image; In this step, the fully connected layer of the preset spatial spectral autoencoder projects the feature vector at each position, obtaining a flattened feature sequence equal to the number of bands. The feature sequence is then reconstructed to obtain the spatial spectral features corresponding to the target image. Specifically, the fully connected layer of the preset spatial spectral autoencoder projects the feature vector at each position from... The first dimension is projected to S dimensions, where S = C / G is the number of bands contained in each spectral group. The flattened sequence is then reassembled into spatial-spectral groups, the dimensional order is adjusted, and all spectral groups are restored to their original input dimensions. Meanwhile, to obtain a global feature representation at the sample level, average pooling is performed on the encoder output along the sequence dimension: features = encoded.mean(dim=1).

[0055] Step S205: Using the pixels of the target image as graph nodes, the spatial spectral features as node features of the pixels, and the connecting lines of the neighboring pixels corresponding to the pixels obtained by cosine similarity calculation as edges, a graph data structure of the target image is constructed. In the specific implementation process, this step defines the pixels of the target image. Graph nodes for constructing the graph data structure; define the spatial spectral features. These are the node features of the graph nodes; therefore, the graph node feature matrix is: . No. i The spatial spectral features of a pixel are defined as follows: ; Calculate the cosine similarity between each pixel; the mathematical formula is shown below:

[0056] in, The spatial spectral feature of the i-th pixel; The spatial spectral feature of the j-th pixel; For the first i The pixel and the first j Cosine similarity of each pixel.

[0057] The neighboring nodes of the pixel are filtered based on the cosine similarity of each pixel to obtain the sparse adjacency matrix corresponding to the pixel; the mathematical expression of the sparse adjacency matrix is ​​as follows:

[0058] in, For the first i A sparse adjacency matrix of pixels; For pixels i of K The set of neighboring nodes is selected by retaining only those nodes with the highest cosine similarity to the given node. k Each neighboring pixel.

[0059] The sparse adjacency matrix is ​​symmetrically normalized to construct the graph data structure of the target image. The mathematical expression is as follows:

[0060] Wherein, the degree matrix is ; The adjacency matrix is ​​the symmetric normalized adjacency matrix; Let be the degree matrix, representing the sum of the connection weights of node i; This is the original coefficient matrix.

[0061] Step S206: Based on the graph data structure, a neighborhood aggregation method is used to perform graph convolution calculation to obtain target node feature vectors with enhanced spatial context, connectivity, and boundary integrity. In the specific implementation process of this step, the mathematical expression for graph convolution calculation is as follows:

[0062] in, The feature vector of the target node; It is the feature matrix of the node, which represents the spatial spectral features of the pixel; This is the normalized adjacency matrix; The weight matrix is ​​a learnable matrix; For bias; The mathematical formula for calculating the feature transformation is:

[0063] The mathematical formula for neighborhood propagation is:

[0064] in, For learnable weight matrix, This is a bias term.

[0065] Step S207: The preset graph convolutional network classifier is used to classify the feature vector of the target node to obtain the pixel probability value of the target image, so as to obtain the camouflage target recognition result of the UAV image of the target to be camouflaged.

[0066] In this step, the graph structure constructed above is used to achieve hierarchical feature learning and pixel classification through multi-layer graph convolution. First, the output features of the SGF autoencoder are used to construct a graph structure, and then a normalization operation is performed. The mathematical expression is as follows:

[0067] in, The initial node features are obtained after L2 normalization; It is a continuous set of content, which is L2 normalization, and its purpose is to scale the feature vector to a unit length.

[0068] For the normalized features, end-to-end feature learning is achieved using multi-layer graph convolution operations to obtain the camouflage target recognition result of the UAV image to be camouflaged. Layer: The mathematical formula for calculating the pixel probability value of the target image is as follows:

[0069] in, forl Pixel probability values ​​of the layer; This is the activation function.

[0070] This application solves the detection failure problem caused by spectral similarity by calculating the spectral difference between the target and the background, automatically selecting the 708–900nm near-infrared discriminative band, and eliminating easily confused bands such as visible light. It employs a spatial-spectral Transformer encoder to simultaneously model local spatial context, spectral grouping correlations, and long-range dependencies, capturing the spatial-spectral features of camouflaged targets more comprehensively than traditional CNNs, avoiding feature loss and local overfitting. Unsupervised training using autoencoder reconstruction loss eliminates the need for extensive pixel-level annotation, significantly lowering the data threshold for engineering applications while improving feature robustness. Graph Convolutional Networks (GCNs) are used to explicitly model semantic relationships between pixels, effectively improving issues such as target holes, blurred boundaries, and background misjudgment, achieving pixel-level accurate classification. This application, through a joint scheme of discriminative band selection, spatial-spectral Transformer encoding, and graph-convolutional integral classification, effectively addresses the problems of high spectral similarity between camouflaged targets and the background, insufficient utilization of spatial-spectral features, and inadequate contextual modeling, achieving stable detection of camouflaged targets with high accuracy, high completeness, and low annotation cost in UAV hyperspectral imagery.

[0071] Another embodiment of this application provides a camouflaged target identification device 300 based on UAV imagery, such as... Figure 3 As shown, it includes: The identification module 301 is used to identify the image of the UAV to be disguised target and obtain the target image of the UAV to be disguised target in the near-infrared band. Feature extraction module 302 is used to extract features from the target image using a preset spatial spectral autoencoder to obtain spatial spectral features corresponding to the target image; The classification module 303 is used to classify the pixels of the target image based on the spatial spectral features using a preset graph convolutional network classifier, so as to obtain the camouflage target recognition result of the UAV image to be camouflaged.

[0072] In the specific implementation process, the identification module 301 is specifically used to identify the image of the UAV to be disguised as a target, and obtain the target area and the background area; calculate the average reflectance of the target area and the background area in different bands respectively, and obtain the first average reflectance of the target area and the second average reflectance of the background area in different bands; perform calculation processing based on the first average reflectance and the second average reflectance of the same band to obtain the spectral difference value of the same band; determine the target near-infrared band for identification of the UAV image to be disguised as a target based on the spectral difference value of different bands; extract the image corresponding to the target near-infrared band in the image of the UAV to be disguised as a target, and obtain the target image.

[0073] In the specific implementation process, the feature extraction module 302 is specifically used to: perform spatial spectral encoding on the target image to obtain spectral groups corresponding to each pixel of the target image; encode the joint spatial and spectral features of the local regions of each spectral group to obtain global spatial and spectral encoded features; and decode the global spatial and spectral encoded features to obtain the spatial and spectral features corresponding to the target image.

[0074] In the specific implementation process, the feature extraction module 302 is further used to: perform vector transformation on the spectral groups using the group embedding layer of the preset spatial spectral autoencoder to obtain a first feature vector corresponding to each spectral group; perform position encoding on each of the first feature vectors to obtain a second feature vector corresponding to each of the spectral groups; perform feature extraction on the second feature vectors using the Transformer encoding layer of the preset spatial spectral autoencoder to obtain a third feature vector corresponding to each of the spectral groups; and perform feature fusion and pooling processing on each of the third feature vectors to obtain the global spatial spectral encoded features; wherein, the position encoding includes spatial position encoding and spectral group position encoding.

[0075] In the specific implementation process, the feature extraction module 302 is further used to: extract features from the second feature vector using a multi-head self-attention network of the Transformer coding layer to obtain a fourth feature vector; perform residual linking and layer normalization processing based on the second feature vector and the fourth feature vector to obtain a fifth feature vector; extract features from the fifth feature vector using a feedforward network of the Transformer coding layer to obtain a sixth feature vector; and perform residual linking and layer normalization processing based on the fourth feature vector and the sixth feature vector to obtain a third feature vector corresponding to each of the spectral groups.

[0076] In the specific implementation process, the classification module 303 is specifically used to construct a graph data structure of the target image, using the pixels of the target image as graph nodes, the spatial spectral features as node features of the pixels, and the connection lines of the neighboring pixels corresponding to the pixels obtained by cosine similarity calculation as edges; based on the graph data structure, a neighborhood aggregation method is used to perform graph convolution calculation processing to obtain target node feature vectors with enhanced spatial context, connectivity, and boundary integrity; the preset graph convolutional network classifier is used to classify the target node feature vectors to obtain the pixel probability values ​​of the target image, so as to obtain the camouflage target recognition result of the UAV image to be camouflaged.

[0077] In the specific implementation process, the classification module 303 is also used to: define the pixels of the target image as graph nodes to be constructed in the graph data structure; define the spatial spectral features as node features of the graph nodes; calculate the cosine similarity between each pixel; filter the neighboring nodes of the pixel based on each cosine similarity to obtain a sparse adjacency matrix corresponding to the pixel; and perform symmetric normalization processing on the sparse adjacency matrix to construct the graph data structure of the target image.

[0078] This application effectively addresses the problems of high spectral similarity between camouflaged targets and the background, insufficient utilization of spatial spectral features, and lack of contextual modeling by employing a joint scheme of discriminative band selection, spatial-spectral Transformer encoding, and image-volume integral class. This enables stable detection of camouflaged targets with high precision, high integrity, and low annotation cost under UAV hyperspectral imagery.

[0079] The technical effects of this invention will be explained below with reference to experimental data: Based on noise-free data filtered by bands, comparisons of various methods on a self-built camouflaged target dataset are shown below. Figure 4As shown, (a) is the original input image; (b) is the true label; (c) is the result of the minimum distance classifier (MDC) method; (d) is the result obtained using the support vector machine (SVM) method; (e) is the result obtained using the convolutional neural network (CNN) method; (f) is the result obtained using the hybrid spectral spatial network (HyBridSN) method; and (g) is the result obtained using the method of this application. Methods such as MDC based on manually designed spectral angles and SVM based on hyperplanes perform poorly in classification when the target and background spectra are highly similar. Specifically, this manifests as misclassification of the background (e.g., grass, trees) and the presence of numerous holes within the target region. This is mainly because these methods rely solely on the spectral information of individual pixels and fail to effectively utilize their spatial context. Although deep learning methods such as CNN and HybridSN have been proven effective in many hyperspectral classification tasks, they still exhibit slight misjudgments of the background in the camouflage scenario of this study, and noise is still present in the classification results. This indicates that while these methods possess some spatial feature extraction capabilities, their ability to model complex spatial structures and global contextual relationships remains insufficient. Table 1 shows the quantitative indicators for each comparison method: Table 1: Quantitative Indicators of Each Comparison Method

[0080] The quantitative metrics for each comparative method are shown in Table 1. The four metrics—OA, IoU, F1 score, and Kappa coefficient—are largely consistent with the conclusions of the qualitative interpretation above. MDC, as a linear model based on handcrafted features, exhibits the lowest performance across all metrics, revealing the inherent limitations of traditional methods in extracting complex hyperspectral features. SVM's performance (OA: 90.72%) is significantly improved compared to MDC, demonstrating that it addresses the nonlinear classification problem in high-dimensional feature spaces to some extent through kernel functions. However, its performance in IoU (39.90%) and F1 score (57.04%) is mediocre, indicating insufficient ability to classify and process pixel details at target boundaries. CNN and HybridSN models achieved excellent performance (OA of 96.50% and 95.63%, respectively), fully demonstrating the nonlinear learning capabilities of deep learning models. Their significant improvements in IoU and F1 score further indicate that convolutional operations can effectively capture local spectral context information between pixels.

[0081] In comparison, the proposed method achieves state-of-the-art results across all evaluation metrics. Specifically, the operational accuracy (OA) reaches 99.51%, and the Kappa coefficient reaches 96.90%, indicating that the model's predictions are almost perfectly consistent with the true labels. Notably, the method achieves IoU and F1 scores of 94.49% and 97.17%, respectively. This result strongly demonstrates the effectiveness of the proposed method: the framework eliminates redundant information through band filtering, allowing the model to focus on the most discriminative bands; it effectively combines spatial-spectral information and models long-distance dependencies between bands through spatial-spectral feature encoding and a Transformer encoder, thereby reconstructing more discriminative feature representations; finally, it explicitly models the spatial context relationships between pixels in the image using a GCN classifier, achieving accurate detection of camouflaged targets.

[0082] Another embodiment of this application provides a storage medium storing a computer program, which, when executed by a processor, implements the following method steps: Step 1: Identify the image of the UAV image to be disguised to obtain the target image in the near-infrared band of the UAV image to be disguised; Step 2: Use a preset spatial spectral autoencoder to extract features from the target image to obtain the spatial spectral features corresponding to the target image; Step 3: Based on the spatial spectral features, a preset graph convolutional network classifier is used to classify the pixels of the target image to obtain the camouflage target recognition result of the UAV image to be camouflaged.

[0083] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0084] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0085] The specific implementation process of the above method steps can be found in any of the above embodiments of the camouflaged target recognition method based on UAV imagery, and will not be repeated here.

[0086] This application effectively addresses the problems of high spectral similarity between camouflaged targets and the background, insufficient utilization of spatial spectral features, and lack of contextual modeling by employing a joint scheme of discriminative band selection, spatial-spectral Transformer encoding, and image-volume integral class. This enables stable detection of camouflaged targets with high precision, high integrity, and low annotation cost under UAV hyperspectral imagery.

[0087] Another embodiment of this application provides an electronic device, which can be a server. The electronic device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the program is executed by the processor, it implements the functions or steps of a server-side method for identifying camouflaged targets based on UAV imagery.

[0088] In one embodiment, an electronic device is provided, which can be a client. The electronic device includes a processor, memory, a network interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the program is executed by the processor, it implements client-side functions or steps of a camouflaged target recognition method based on UAV imagery.

[0089] Another embodiment of this application provides an electronic device, including at least a memory and a processor. The memory stores a computer program, and the processor, when executing the computer program in the memory, performs the following method steps: Step 1: Identify the image of the UAV image to be disguised to obtain the target image in the near-infrared band of the UAV image to be disguised; Step 2: Use a preset spatial spectral autoencoder to extract features from the target image to obtain the spatial spectral features corresponding to the target image; Step 3: Based on the spatial spectral features, a preset graph convolutional network classifier is used to classify the pixels of the target image to obtain the camouflage target recognition result of the UAV image to be camouflaged.

[0090] The specific implementation process of the above method steps can be found in any of the above embodiments of the camouflaged target recognition method based on UAV imagery, and will not be repeated here.

[0091] This application effectively addresses the problems of high spectral similarity between camouflaged targets and the background, insufficient utilization of spatial spectral features, and lack of contextual modeling by employing a joint scheme of discriminative band selection, spatial-spectral Transformer encoding, and image-volume integral class. This enables stable detection of camouflaged targets with high precision, high integrity, and low annotation cost under UAV hyperspectral imagery.

[0092] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. Those skilled in the art can make various modifications or equivalent substitutions to this application within the scope and nature of this application, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A camouflage target identification method based on unmanned aerial vehicle (UAV) images, characterized in that, include: Image recognition is performed on the image of the drone image to be disguised to obtain a target image in the near-infrared band of the drone image to be disguised to be identified. The target image is feature extracted using a preset spatial spectral autoencoder to obtain the spatial spectral features corresponding to the target image; Based on the spatial spectral features, a preset graph convolutional network classifier is used to classify the pixels of the target image to obtain the camouflage target recognition result of the UAV image to be camouflaged; The process of performing image recognition on the UAV image of the target to be camouflaged to obtain a target image in the near-infrared band specifically includes: The image of the drone to be disguised is identified to obtain the target area and the background area; The average reflectance of the target area and the background area in different bands are calculated respectively to obtain the first average reflectance of the target area in different bands and the second average reflectance of the background area in different bands. The spectral difference value of the same band is obtained by calculating the first average reflectance and the second average reflectance based on the same band. The near-infrared band of the target to be identified in the UAV image is determined based on the spectral difference values ​​of different bands. Extract the image corresponding to the near-infrared band of the target from the image of the UAV image for target identification to obtain the target image; The target's near-infrared band is 708nm-900nm; The step of extracting features from the target image using a preset spatial spectral autoencoder to obtain spatial spectral features corresponding to the target image specifically includes: Spatial spectral encoding is performed on the target image to obtain spectral groups corresponding to each pixel of the target image; The joint spatial-spectral features of local regions of each of the aforementioned spectral groups are encoded to obtain global spatial-spectral encoded features; The global spatial spectrum coding features are decoded to obtain the spatial spectrum features corresponding to the target image.

2. The method of claim 1, wherein, The process of encoding the joint spatial-spectral features of local regions of each of the spectral groups to obtain global spatial-spectral encoded features specifically includes: The preset spatial spectral autoencoder is used to perform vector transformation on the spectral groups to obtain a first feature vector corresponding to each spectral group. Position encoding is performed on each of the first feature vectors to obtain a second feature vector corresponding to each of the spectral groups; The second feature vector is extracted using the Transformer encoding layer of the preset spatial spectral autoencoder to obtain a third feature vector corresponding to each of the spectral groups. The third feature vectors are subjected to feature fusion and pooling to obtain the global spatial spectrum coding features; The location coding includes spatial location coding and spectral group location coding.

3. The method of claim 2, wherein, The step of extracting features from the second feature vector using the Transformer encoding layer of the preset spatial spectral autoencoder to obtain a third feature vector corresponding to each of the spectral groups specifically includes: A multi-head self-attention network with a Transformer coding layer is used to extract features from the second feature vector to obtain the fourth feature vector; Based on the second and fourth feature vectors, residual linking and layer normalization are performed to obtain the fifth feature vector; The feedforward network of the Transformer coding layer is used to extract features from the fifth feature vector to obtain the sixth feature vector; Based on the fourth and sixth feature vectors, residual linking and layer normalization are performed to obtain the third feature vector corresponding to each of the spectral groups.

4. The method as described in claim 1, characterized in that, The step of classifying the pixels of the target image using a preset graph convolutional network classifier based on the spatial spectral features to obtain the camouflage target recognition result of the UAV image to be camouflaged specifically includes: The graph data structure of the target image is constructed by using the pixels of the target image as graph nodes, the spatial spectral features as node features of the pixels, and the connection lines of the neighboring pixels corresponding to the pixels obtained by cosine similarity calculation as edges. Based on the graph data structure, a neighborhood aggregation method is used to perform graph convolution calculation to obtain target node feature vectors with enhanced spatial context, connectivity, and boundary integrity. The preset graph convolutional network classifier is used to classify the feature vector of the target node to obtain the pixel probability value of the target image, so as to obtain the camouflage target recognition result of the drone image to be camouflaged.

5. The method as described in claim 4, characterized in that, The process of constructing a graph data structure for the target image, using pixels of the target image as graph nodes, spatial spectral features as node features of the pixels, and connecting lines between neighboring pixels corresponding to the pixel obtained by cosine similarity calculation as edges, specifically includes: The pixels of the target image are defined as graph nodes of the graph data structure to be constructed; The spatial spectral feature is defined as the node feature of the graph node; Calculate the cosine similarity between each of the aforementioned pixels; The neighboring nodes of the pixel are filtered based on the cosine similarity of each pixel to obtain the sparse adjacency matrix corresponding to the pixel. The sparse adjacency matrix is ​​symmetrically normalized to construct the graph data structure of the target image.

6. A camouflage target identification device based on UAV imagery, used to implement the camouflage target identification method based on any one of claims 1 to 5, characterized in that, include: The identification module is used to identify the image of the drone to be disguised and obtain a target image in the near-infrared band of the image of the drone to be disguised. The feature extraction module is used to extract features from the target image using a preset spatial spectral autoencoder to obtain spatial spectral features corresponding to the target image; The classification module is used to classify the pixels of the target image based on the spatial spectral features using a preset graph convolutional network classifier, so as to obtain the camouflage target recognition result of the UAV image to be camouflaged.

7. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the camouflaged target recognition method based on UAV imagery as described in any one of claims 1-5.

8. An electronic device, characterized in that, It includes at least a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program in the memory, implements the steps of the camouflage target recognition method based on UAV imagery as described in any one of claims 1-5.

Images