An end-to-end hyperspectral image multi-classification camouflage target segmentation method and device

Abstract

The present application relates to a kind of end-to-end hyperspectral image multi-classification camouflage target segmentation method and equipment, belong to target detection technical field.The method includes the following steps: constructing the hyperspectral image data set containing camouflage target;Using data set to the initial camouflage target segmentation model of pre-construction is iteratively trained, and the convergence camouflage target segmentation model is obtained;Camouflage target segmentation model is used to the wave band information extraction and feature fusion focus of input image, and based on the feature map after feature fusion focus carries out camouflage target classification prediction;The hyperspectral image to be detected is input into camouflage target segmentation model, and the multi-classification camouflage target segmentation result of identification is obtained.The present application solves the problem that the detection ability and precision of prior art to hyperspectral image camouflage target are not high.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to an end-to-end hyperspectral image multi-classification camouflage target segmentation method and device. Background Technology

[0002] Multi-class camouflage target recognition is a highly challenging visual task, especially for hyperspectral data. Hyperspectral images typically possess dozens or even hundreds of continuous spectral bands within a specific spectral range, offering rich spectral information. This high band resolution provides extremely effective discriminative information for distinguishing different substances and materials, enabling the differentiation of camouflage targets and backgrounds that are difficult to distinguish in the visible light band. However, the high correlation and low difference between adjacent bands present challenges due to the large number of spectral bands, the large data volume, and the abundance of redundant information. Therefore, before processing tasks related to hyperspectral images, dimensionality reduction processing of hyperspectral images is necessary. On the other hand, multi-class camouflage recognition tasks based on hyperspectral images require network models capable of extracting rich and important features from both spatial and spectral ranges. Especially for multi-class camouflage target recognition tasks, the quality of feature extraction is crucial. Summary of the Invention

[0003] Based on the above analysis, the present invention aims to provide an end-to-end hyperspectral image multi-classification camouflage target segmentation method and device to solve the problem of low detection capability and accuracy of existing technologies for camouflage targets in hyperspectral images.

[0004] The objective of this invention is mainly achieved through the following technical solutions:

[0005] This invention provides an end-to-end hyperspectral image multi-class camouflage target segmentation method, comprising:

[0006] Construct a hyperspectral image dataset containing camouflaged targets;

[0007] The pre-constructed initial camouflage target segmentation model is iteratively trained using the dataset to obtain a converged camouflage target segmentation model. The camouflage target segmentation model is used to extract band information and fuse and focus features from the input image, and to perform camouflage target classification and prediction based on the feature map after feature fusion and focusing.

[0008] The hyperspectral image to be detected is input into the camouflage target segmentation model to identify multi-class camouflage target segmentation results.

[0009] Furthermore, the camouflaged target segmentation model includes: a band information extraction network, a backbone network, and a feature optimization network;

[0010] The band information extraction network is used to extract and reduce the dimensionality of the input hyperspectral image;

[0011] The backbone network is used to perform size compression and feature extraction on the feature maps output by the band information extraction network at multiple different scales, obtaining backbone feature maps of multiple different sizes;

[0012] The feature optimization network is used to perform feature fusion and focusing based on the backbone feature maps of different sizes, and through classification prediction, obtain the camouflage target segmentation result.

[0013] Furthermore, the band information extraction network includes a spatial pooling layer, a 1×k convolutional layer, and a 1×1 convolutional layer arranged in sequence; the band information extraction network is used to extract and reduce the dimensionality of the band information of the input hyperspectral image, including:

[0014] Successively pass the input hyperspectral image through the spatial pooling layer and the 1×k convolutional layer to perform pooling operations and convolutional operations for band information extraction;

[0015] Perform corresponding channel multiplication operations on the image obtained by band information extraction and the input hyperspectral image;

[0016] Input the image obtained after multiplication into the 1×1 convolutional layer for linear mixing of channel information, obtaining the output of the band information extraction network.

[0017] Furthermore, n + 1 backbone feature maps are obtained through the backbone network; the feature optimization network includes n feature fusion and focusing modules; among them,

[0018] The nth feature fusion and focusing module is used to perform feature fusion based on the (n + 1)th backbone feature map and the nth backbone feature map;

[0019] The (n - i)th feature fusion and focusing module is used to perform feature fusion on the feature map output by the (n - i + 1)th feature fusion and focusing module and the (n - i)th backbone feature map, where i is an integer with 1 ≤ i < n;

[0020] Based on the feature map output by the first feature fusion and focusing module, perform classification prediction to obtain the camouflage target segmentation result.

[0021] Furthermore, the feature fusion and focusing module includes a deconvolution module, a first channel information fusion module, a second channel information fusion module, a third channel information fusion module, and a spatial information processing module;

[0022] The deconvolution module and the first channel information fusion module are serially arranged, and perform image size amplification and band information extraction and fusion on the high-order feature map in the two input feature maps, obtaining a first band information feature map;

[0023] The second channel information fusion module is used to extract and fuse band information from the low-order feature maps of the two input feature maps to obtain the second band information feature map;

[0024] The first band information feature map and the second band information feature map are added together by the add operation to obtain the band information fusion feature map;

[0025] The third channel information fusion module and the spatial information processing module are set up in parallel, and are respectively used to extract band information and spatial information from the band information fusion feature map to obtain the third band information feature map and the spatial information feature map;

[0026] The third-band information feature map and the spatial information feature map are added together using the add operation to obtain the feature map output by the feature fusion focusing module.

[0027] Furthermore, the first channel information fusion module, the second channel information fusion module, and the third channel information fusion module all include a spatial pooling layer and a 1×k convolutional layer arranged sequentially, which are used to perform spatial pooling and convolution operations on the feature map of the input channel information fusion module in sequence, and multiply the input feature map with the feature map output by the 1×k convolutional layer to obtain the band information feature map.

[0028] Furthermore, the spatial information processing module includes a 1×1 convolutional layer and a sigmoid activation layer, which are used to process the feature map of the input spatial information processing module through linear convolution and activation operations in sequence, and add the feature map output by the activation layer to the feature map of the input spatial information processing module through an add operation to obtain a spatial information feature map.

[0029] Furthermore, the backbone network is a feature compression and extraction network based on ResNet152.

[0030] Furthermore, the construction of the hyperspectral camouflage target dataset includes:

[0031] Acquire hyperspectral images containing camouflaged targets;

[0032] The dimensions of the hyperspectral image are cropped to h×w×c, where h, w, and c are the height, width, and number of spectral channels of the image, respectively.

[0033] A three-channel pseudo-color image is generated for each hyperspectral image. The camouflaged target is labeled according to the segmentation task based on the pseudo-color image. A hyperspectral camouflaged target dataset is constructed based on the labeled pseudo-color image.

[0034] On the other hand, an electronic device is also provided, including at least one processor and at least one memory communicatively connected to said processor;

[0035] The memory stores instructions that can be executed by the processor to implement the aforementioned end-to-end hyperspectral image multi-class camouflage target segmentation method.

[0036] The beneficial effects of this technical solution are as follows:

[0037] 1. The camouflage target segmentation model constructed and trained in this invention improves the accuracy of camouflage target segmentation by extracting band information and fusing and focusing features from hyperspectral images, making full use of the joint information of hyperspectral image channels and space to optimize feature map quality;

[0038] 2. By using a pre-existing band information extraction network, not only can image dimensionality reduction and key band extraction be achieved, but end-to-end integrated segmentation of hyperspectral images can also be realized. Furthermore, by using a feature optimization network, the spatial band feature fusion and focusing capabilities can be improved, thereby enabling multi-class camouflage target recognition in hyperspectral images. This method does not require the addition of additional equipment and facilities and can efficiently and conveniently achieve multi-class camouflage target segmentation in hyperspectral images.

[0039] 3. This invention, by setting a 1×k convolutional layer and mixing information from k adjacent spectral bands, achieves the extraction of effective information from similar bands, enabling the model to have cross-image reasoning capabilities. It overcomes the limitation of hyperspectral classification tasks that rely on training and reasoning based on a single hyperspectral image, and also overcomes the limitation of hyperspectral image anomaly detection that relies on binary target recognition based on a single image that only distinguishes between foreground and background.

[0040] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0041] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0042] Figure 1 This is a schematic diagram of an end-to-end hyperspectral image multi-class camouflage target segmentation method according to an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the band information extraction module structure according to an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the channel information fusion module according to an embodiment of the present invention;

[0045] Figure 4These are different visual effect diagrams formed after the fusion of different weighted vector channels in an embodiment of the present invention.

[0046] Figure 5 Different visual effects formed by spatial focusing through different spatial weights in embodiments of the present invention;

[0047] Figure 6 These are different visual effect diagrams formed by the corresponding superimposed channel weights and spatial weights in embodiments of the present invention, based on the fusion focusing module.

[0048] Figure 7 This is a test result of hyperspectral image data according to an embodiment of the present invention. Detailed Implementation

[0049] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0050] A specific embodiment of the present invention, such as Figure 1 As shown, an end-to-end hyperspectral image multi-class camouflage target segmentation method is disclosed, including the following steps:

[0051] Step S1: Construct a hyperspectral image dataset containing camouflaged targets;

[0052] Specifically, hyperspectral (multispectral) images can be acquired using a hyperspectral (multispectral) camera. The acquisition method can be either line scanning or snapshot-style. For images acquired by the hyperspectral (multispectral) camera, the image dimensions are first uniformly cropped to h×w×c, where h, w, and c are the image height, width, and number of spectral channels, respectively. For example, in this embodiment, the acquired images are uniformly cropped to a size of 960×1024×180, where 960, 1024, and 180 are the image height, width, and number of spectral channels, respectively. Further, for each hyperspectral image with a camouflaged target, a three-channel pseudo-color image is generated. Based on the pseudo-color image, the acquired images are labeled with camouflaged targets according to the requirements of the multi-class camouflaged target segmentation task. A hyperspectral image dataset containing camouflaged targets is then constructed based on the labeled images. In practical applications, the labeled images can be randomly divided into a training set, a validation set, and a test set; in this embodiment, the division is 8:1:1.

[0053] S2, using the dataset, iteratively trains the pre-built initial camouflage target segmentation model to obtain a converged camouflage target segmentation model; the camouflage target segmentation model is used to extract band information and fuse and focus features from the input image, and performs camouflage target classification prediction based on the feature map after feature fusion and focusing;

[0054] Specifically, the camouflage target segmentation model includes: a band information extraction network, a backbone network, and a feature optimization network; among which,

[0055] The band information extraction network comprises a spatial pooling layer, a 1×k convolutional layer, and a 1×1 convolutional layer arranged sequentially, used to extract and reduce the dimensionality of the input hyperspectral image, including:

[0056] The input hyperspectral image is sequentially passed through a spatial pooling layer and a 1×k convolutional layer for pooling and convolution operations to extract band information. The image obtained by extracting band information is then multiplied with the input hyperspectral image for corresponding channels. The resulting image is then input into a 1×1 convolutional layer for linear mixing of channel information to obtain the output of the band information extraction network.

[0057] In this embodiment, the input hyperspectral image, after passing through the band information extraction module, enables the interaction and fusion of information between different bands, ultimately outputting a fused three-channel feature map. More specifically, such as Figure 2 As shown, the input hyperspectral image X, with dimensions w×h×c, undergoes one spatial pooling and one 1×k convolution to obtain a one-dimensional vector 1×1×c. This one-dimensional vector is then multiplied by the corresponding channel of the input hyperspectral image to obtain a new hyperspectral image, still with dimensions w×h×c. Finally, a 1×1 convolution operation with 180 input channels and 3 output channels outputs a combined three-channel feature map with dimensions w×h×3. The 1×k convolution in the band information extraction module aims to mix the information of k adjacent bands, where k can be 3, 5, or 7; in this embodiment, it is set to 5. This approach is inspired by the continuous bands and gradual wavelength changes in hyperspectral images. The 1×k convolution achieves the extraction of effective information from similar bands; the final 1×1 convolution layer achieves linear mixing of channel information while preserving the original spatial information. The band information extraction module proposed in this invention achieves effective reduction of the spectral dimension of hyperspectral images through a learnable method, while learning as many optimal image feature representations of the original hyperspectral images as possible that are effective for the task. This improves the accuracy of camouflage target segmentation and plays a crucial role in the design and implementation of camouflage target segmentation models.

[0058] The backbone network is used to perform size compression and feature extraction on the feature map output by the band information extraction network at multiple different scales, resulting in backbone feature maps of multiple different sizes.

[0059] Specifically, in this embodiment, the backbone network is a feature compression and extraction network based on ResNet152, which is used to extract features at multiple scales from the image output by the band information extraction network. Exemplarily, in this embodiment, features are extracted in five stages through the backbone network. In each stage, the spatial resolution is reduced (both the width and height are reduced by half), and at the same time, the number of channels is increased. The sizes of the output feature maps after the third, fourth, and fifth stages are 1 / 8, 1 / 16, and 1 / 32 of the original image respectively, obtaining three backbone feature maps with different scales. Specifically, the size of the image after dimensionality reduction by the band information extraction network is 960×1024×3. After the first and second stage size compression and feature extraction by the backbone network, feature maps with feature resolutions of 480×512*64 and 240×256*128 are obtained. The feature map extracted in the second stage is subjected to the third, fourth, and fifth stage feature extractions, and the resolution is reduced by half in turn, obtaining the first backbone feature map, the second backbone feature map, and the third backbone feature map with resolutions of 120*128*256, 60×64*512, and 30×32*1024 respectively.

[0060] The feature optimization network is used to perform feature fusion and focusing based on the backbone feature maps of different sizes, and through classification prediction, obtain the camouflage target segmentation result.

[0061] Specifically, n + 1 backbone feature maps can be obtained through the backbone network; the feature optimization network includes n feature fusion and focusing modules; among them, the nth feature fusion and focusing module is used to perform feature fusion based on the (n + 1)th backbone feature map and the nth backbone feature map; the (n - i)th feature fusion and focusing module is used to perform feature fusion on the feature map output by the (n - i + 1)th feature fusion and focusing module and the (n - i)th backbone feature map, where i is an integer of 1 ≤ i < n; the feature map output by the first feature fusion and focusing module is used as the target feature map for classification prediction to obtain the camouflage target segmentation result.

[0062] Specifically, the feature fusion and focusing module includes a transposed convolution module, a first channel information fusion module, a second channel information fusion module, a third channel information fusion module, and a spatial information processing module; among them,

[0063] The transposed convolution module and the first channel information fusion module are serially arranged to perform image size amplification and band information extraction and fusion on the high-order feature map in the two input feature maps, obtaining the first band information feature map;

[0064] The second channel information fusion module is used to perform band information extraction and fusion on the low-order feature map in the two input feature maps, obtaining the second band information feature map;

[0065] The first band information feature map and the second band information feature map are added together by the add operation to obtain the band information fusion feature map;

[0066] The third-channel information fusion module and the spatial information processing module are set up in parallel, and are used to extract band information and spatial information from the band information fusion feature map, respectively, to obtain the third-band information feature map and the spatial information feature map;

[0067] The third-band information feature map and the spatial information feature map are added together using the add operation to obtain the feature map output by the feature fusion focusing module.

[0068] The first channel information fusion module, the second channel information fusion module, and the third channel information fusion module all include a spatial pooling layer and a 1×k convolutional layer arranged sequentially. These layers are used to perform spatial pooling and convolution operations on the feature maps of the input channel information fusion modules in sequence, and then multiply the input feature map with the feature map output by the 1×k convolutional layer to obtain the band information feature map.

[0069] The spatial information processing module includes a 1×1 convolutional layer and a sigmoid activation layer. It processes the feature map of the input spatial information processing module through linear convolution and activation operations in sequence. The feature map output by the activation layer is added to the feature map of the input spatial information processing module through an add operation to obtain the spatial information feature map.

[0070] In this embodiment, feature maps of different sizes are fused by the channel information fusion module. This achieves resolution alignment between adjacent feature maps while efficiently focusing on key feature regions on the feature maps, and simultaneously enables effective fusion of higher-order and lower-order features. For example, as shown... Figure 3 As shown, adjacent-order feature maps of the hyperspectral image are simultaneously input into the channel information fusion module. The higher-order feature map is first augmented by a factor of 2 through a single deconvolution. The augmented feature map and the lower-order feature map are processed in parallel, each undergoing spatial pooling and a 1×k convolution to obtain a corresponding one-dimensional vector 1×1×c. Each one-dimensional vector is then multiplied by the corresponding channel of the input feature map to obtain a new feature map, still with dimensions w×h×c. The two new feature maps are then converted to have the same number of channels through a 1×1 convolution and added together using an add operation. After the add operation, the dimension of the feature map remains unchanged, but the information content in each dimension increases, providing beneficial feature maps for the final camouflage target segmentation. The resulting feature map is then processed in parallel by the channel information fusion module and the spatial information processing module to achieve channel information fusion and spatial information processing, resulting in a feature map that simultaneously achieves the fusion and focusing of important features in both spatial and channel dimensions.

[0071] Specifically, Figure 4Different visual effects are formed after channel fusion with different weight vectors. Obviously, different weight vector fusions have different effects on highlighting camouflaged targets. Figure 5 Different spatial weights result in different visual effects of spatial focusing. Similarly, spatial weights also play a crucial role in highlighting camouflaged targets. Figure 6 These are different visual effect images generated by the fusion focusing module, corresponding to the superimposed channel weights and spatial weights. Therefore, the feature fusion focusing module can effectively achieve the fusion focusing of image features, thus facilitating the exposure of camouflaged targets.

[0072] In this embodiment, the feature optimization network achieves feature fusion of feature maps of different sizes by repeatedly calling the channel information fusion module. For example, the feature optimization network first uses a channel information fusion module to fuse and focus the fifth-order features (third backbone feature map) and the fourth-order features (second backbone feature map). Then, it fuses and focuses the feature map obtained after the first fusion and focusing with the third-order features (first backbone feature map) to obtain a target feature map for camouflage target classification prediction. The first call to the channel information fusion module outputs a feature map with a size of 1 / 16 of the original image; the second call outputs a feature map with a size of 1 / 8 of the original image. Then, a deconvolution operation is used to upsample the 1 / 8 size feature map by eight times, resulting in a target feature map with the same size as the original image, whose number of channels equals the number of camouflage target categories plus the number of background categories. The network is trained using training and validation datasets, and through iterative optimization, a converged camouflage target segmentation model is obtained. The trained camouflage target segmentation model is then used to detect and identify the test dataset. When the model provides the target segmentation result, the result cannot be directly visualized because the target feature map has only 7 channels. Therefore, the prediction results are first mapped to a single-channel prediction category map based on the highest probability, and then fixedly mapped to a color palette based on the prediction category map to visualize the detection results. For example... Figure 7 The image shows the test results for a hyperspectral image, where (a) is the pseudo-color image of the hyperspectral image, (b) is the corresponding label image, and (c) is the visualization of the prediction results of the model.

[0073] S3, input the hyperspectral image to be detected into the camouflage target segmentation model, and identify the multi-class camouflage target segmentation results;

[0074] In practical applications, real-time acquired hyperspectral images are input into a trained camouflage target segmentation model. This model fully utilizes the spectral and spatial information of the hyperspectral images to efficiently complete multi-class camouflage target segmentation without the need for additional detection equipment or devices. This achieves accurate identification of multi-class camouflage targets and improves the adaptability and practicality of the camouflage target segmentation method to various application scenarios.

[0075] In summary, this invention discloses an end-to-end hyperspectral image multi-class camouflage target segmentation method and device. The camouflage target segmentation model constructed and trained by this method improves the accuracy of camouflage target segmentation by extracting band information and fusing and focusing features from hyperspectral images, fully utilizing the joint information of hyperspectral image channels and space to optimize feature map quality. A pre-processed band information extraction network achieves dimensionality reduction of hyperspectral images while extracting key spectral information, and a feature optimization network enhances the spatial band feature fusion and focusing capability, thereby realizing multi-class camouflage target recognition in hyperspectral images. This method does not require additional equipment or facilities and can efficiently and conveniently achieve multi-class camouflage target segmentation in hyperspectral images. Furthermore, by setting a channel information fusion module containing a 1×k convolutional layer, this invention mixes information from k adjacent spectral bands, achieving the extraction of effective information from similar bands, enabling the model to have cross-image inference capabilities. This overcomes the limitation of hyperspectral classification tasks based on training and inference from a single hyperspectral image, and also overcomes the limitation of hyperspectral image anomaly detection based on binary target recognition that only distinguishes foreground and background from a single image.

[0076] Another embodiment of the present invention also provides an electronic device, including at least one processor and at least one memory communicatively connected to the processor;

[0077] The memory stores instructions that can be executed by the processor to implement the aforementioned end-to-end hyperspectral image multi-class camouflage target segmentation method.

[0078] Those skilled in the art will understand that all or part of the methods in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0079] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An end-to-end hyperspectral image multi-class camouflage target segmentation method, characterized in that, include: Construct a hyperspectral image dataset containing camouflaged targets; The pre-constructed initial camouflage target segmentation model is iteratively trained using the dataset to obtain a converged camouflage target segmentation model. The camouflage target segmentation model is used to extract band information and fuse and focus features from the input image, and to perform camouflage target classification and prediction based on the feature map after feature fusion and focusing. The camouflage target segmentation model includes: a band information extraction network, a backbone network, and a feature optimization network; the band information extraction network is used to extract and reduce the dimensionality of the input hyperspectral image; the backbone network is used to perform size compression and feature extraction on the feature map output by the band information extraction network at multiple different scales to obtain multiple backbone feature maps of different sizes; the feature optimization network is used to perform feature fusion and focusing based on the backbone feature maps of different sizes through a feature fusion and focusing module, and then perform classification prediction to obtain the camouflage target segmentation result; The feature fusion focusing module includes a deconvolution module, a first channel information fusion module, a second channel information fusion module, a third channel information fusion module, and a spatial information processing module. The deconvolution module and the first channel information fusion module are configured serially to perform image size augmentation and band information extraction and fusion on the higher-order feature maps of the two input feature maps, resulting in a first band information feature map. The second channel information fusion module is used to perform band information extraction and fusion on the lower-order feature maps of the two input feature maps, resulting in a second band information feature map. The first and second band information feature maps are then added together using an add operation to obtain a band information fused feature map. The third channel information fusion module and the spatial information processing module are configured in parallel to perform band information extraction and spatial information extraction on the band information fused feature map, respectively, to obtain a third band information feature map and a spatial information feature map. The third band information feature map and the spatial information feature map are then added together using an add operation to obtain the feature map output by the feature fusion focusing module. The hyperspectral image to be detected is input into the camouflage target segmentation model to identify multi-class camouflage target segmentation results.

2. The end-to-end hyperspectral image multi-class camouflage target segmentation method according to claim 1, characterized in that, The band information extraction network includes a spatial pooling layer, a 1×k convolutional layer, and a 1×1 convolutional layer arranged sequentially; the band information extraction network is used to extract and reduce the dimensionality of the input hyperspectral image, including: The input hyperspectral image is sequentially passed through a spatial pooling layer and a 1×k convolutional layer to perform pooling and convolution operations to extract band information. The image obtained by extracting the band information is multiplied by the corresponding channel of the input hyperspectral image; The image obtained after multiplication is input into the 1×1 convolutional layer for linear mixing of channel information to obtain the output of the band information extraction network.

3. The end-to-end hyperspectral image multi-class camouflage target segmentation method according to claim 1, characterized in that, The backbone network yields n+1 backbone feature maps; the feature optimization network comprises n feature fusion and focusing modules; wherein... The nth feature fusion focusing module is used to perform feature fusion based on the (n+1)th backbone feature map and the nth backbone feature map; The (n - i)-th feature fusion focusing module is used to perform feature fusion on the feature map output by the (n - i + 1)-th feature fusion focusing module and the (n - i)-th backbone feature map, where i is an integer satisfying 1 ≤ i < n; Based on the feature map output by the 1st feature fusion focusing module, classification prediction is performed to obtain the segmentation result of the camouflage target.

4. The end-to-end hyperspectral image multi-class camouflage target segmentation method according to claim 1, characterized in that, The first channel information fusion module, the second channel information fusion module, and the third channel information fusion module each include a spatial pooling layer and a 1×k convolutional layer arranged in sequence, which are used to perform spatial pooling operations and convolutional operations on the feature map input to the channel information fusion module in sequence, and multiply the input feature map by the feature map output by the 1×k convolutional layer to obtain a band information feature map.

5. The end-to-end hyperspectral image multi-class camouflage target segmentation method according to claim 1, characterized in that, The spatial information processing module includes a 1×1 convolutional layer and a sigmoid activation layer, which are used to process the spatial information of the feature map input to the spatial information processing module through linear convolutional operations and activation operations in sequence, and add the feature map output by the activation layer to the feature map input to the spatial information processing module through an add operation to obtain a spatial information feature map.

6. The end-to-end hyperspectral image multi-class camouflage target segmentation method according to claim 1, characterized in that, The backbone network is a feature compression and extraction network based on ResNet152.

7. The end-to-end hyperspectral image multi-class camouflage target segmentation method according to claim 1, characterized in that, The construction of the hyperspectral camouflage target dataset includes: Collecting hyperspectral images containing camouflage targets; The dimensions of the hyperspectral image are cropped to h. w c, where h, w, and c are the height, width, and number of spectral channels of the image, respectively; Generating a three-channel false color map corresponding to each hyperspectral image, annotating the camouflage target based on the false color map according to the segmentation task, and constructing a hyperspectral image dataset containing the camouflage target based on the annotated false color map.

8. An electronic device, characterized in that, Comprising at least one processor and at least one memory communicatively connected to the processor; The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the end-to-end hyperspectral image multi-classification camouflage target segmentation method according to any one of claims 1-7.

Images