Hyperspectral Image Feature Processing Method, Classification Method, Apparatus, System and Storage Medium

Abstract

The present application discloses a hyperspectral image feature processing method, a classification method, a device, a system and a storage medium. The hyperspectral image feature processing method first receives a hyperspectral image, then inputs the spatial-spectral domain information of the hyperspectral image into N feature extraction modules respectively to obtain at least N levels of feature maps correspondingly, and then performs fusion processing on the at least N levels of feature maps to obtain a fused feature map, where N is a positive integer greater than or equal to 2, and the at least N levels of feature maps have spatial-spectral features of different scales. Compared with the spatial-spectral feature map of a single scale, the discriminative ability of the fused feature map that fuses multiple spatial-spectral features of different scales in the present application is more prominent, which is beneficial to improving the accuracy of subsequent classification of hyperspectral images.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and particularly relates to a hyperspectral image feature processing method, classification method, device, system and storage medium. Background Art

[0002] Compared with traditional RGB images, hyperspectral images contain richer ground object spatial information and spectral information, and have been widely used in many aspects. For example, in the agricultural field, hyperspectral images can monitor the growth of crops and estimate crop yields; in the military field, hyperspectral images can be used for military target reconnaissance and camouflage recognition through different spectral characteristics; in the field of geological exploration, fine identification, mapping, and exploration of mineral resources can be carried out according to the different spectral properties of different minerals.

[0003] However, the labeled samples of hyperspectral images are relatively scarce. By adopting small sample learning algorithms, the problem of scarce labeled samples of hyperspectral images can be alleviated to a certain extent. However, current small sample learning algorithms usually use features of a single scale for prediction, ignoring features of other scales, resulting in a low accuracy rate for subsequent hyperspectral image classification. Summary of the Invention

[0004] This application aims to at least solve one of the technical problems existing in the prior art. For this purpose, this application provides a hyperspectral image feature processing method, classification method, device, system and storage medium, which can obtain a fused feature map that fuses empty-spectrum features of multiple different scales, and has stronger discriminative ability compared with a single-scale empty-spectrum feature map.

[0005] The first aspect embodiment of this application provides a hyperspectral image feature processing method, including:

[0006] Receiving a hyperspectral image;

[0007] Respectively inputting the empty-spectrum domain information of the hyperspectral image into N feature extraction modules to correspondingly obtain at least N levels of feature maps; where N is a positive integer greater than or equal to 2, and the at least N levels of feature maps have empty-spectrum features of different scales;

[0008] Performing fusion processing on the at least N levels of feature maps to obtain a fused feature map.

[0009] The hyperspectral image feature processing method according to the embodiments of the first aspect of the present application has at least the following beneficial effects: The hyperspectral image feature processing method of the embodiments of the present application first receives a hyperspectral image, and then inputs the spatial-spectral domain information of the hyperspectral image into N feature extraction modules respectively to obtain at least N levels of feature maps correspondingly. Then, the at least N levels of feature maps are fused to obtain a fused feature map, where N is a positive integer greater than or equal to 2, and the at least N levels of feature maps have spatial-spectral features of different scales. Compared with the spatial-spectral feature map of a single scale, the discriminative ability of the fused feature map that fuses multiple spatial-spectral features of different scales in the present application is more prominent, which is beneficial to improving the accuracy of subsequent hyperspectral image classification.

[0010] According to some embodiments of the first aspect of the present application, the feature extraction module includes at least one convolutional block and at least one attention block;

[0011] The step of inputting the spatial-spectral domain information of the hyperspectral image into N feature extraction modules respectively to obtain at least N levels of feature maps correspondingly includes:

[0012] Input the spatial-spectral domain information of the hyperspectral image into the S-th feature extraction module; where S = 1, 2.....N;

[0013] Perform convolution processing on the hyperspectral image through the convolutional block, and perform update processing on the hyperspectral image using attention weights through the attention block to obtain the S-th level of feature map.

[0014] According to some embodiments of the first aspect of the present application, the S-th feature extraction module includes S + 1 convolutional blocks and S attention blocks;

[0015] The step of performing convolution processing on the hyperspectral image through the convolutional block and performing update processing on the hyperspectral image using attention weights through the attention block to obtain the S-th level of feature map includes:

[0016] Perform the first convolution processing on the hyperspectral image through the first convolutional block to obtain the result of the first convolution processing; where the scale of the spatial-spectral features of the result of the first convolution processing is half of the scale of the spatial-spectral domain information of the hyperspectral image;

[0017] Perform the L-th attention weight update processing on the result of the L-th convolution processing through the L-th attention block to obtain the feature map after the L-th attention weight update processing; where L = 1, 2,.....S;

[0018] Perform the (L + 1)-th convolution process on the feature map after the L-th attention weight update through the (L + 1)-th convolution block to obtain the result of the (L + 1)-th convolution process; wherein, the scale of the empty spectral feature of the result of the (L + 1)-th convolution process is half of the scale of the empty spectral feature of the feature map after the L-th attention weight update process.

[0019] Until L = S, take the result of the (L + 1)-th convolution process as the S-th level feature map.

[0020] According to some embodiments of the first aspect of the present application, the process of performing the L-th attention weight update process on the result of the L-th convolution process through the L-th attention block to obtain the feature map after the L-th attention weight update process includes:

[0021] Divide the result of the L-th convolution process into several groups of feature data in the channel dimension;

[0022] Divide each group of the feature data into first feature data and second feature data;

[0023] Obtain the channel attention weight according to the first feature data;

[0024] Obtain the spatial attention weight according to the second feature data;

[0025] Obtain the feature map after the L-th attention weight update process according to the channel attention weight and the spatial attention weight.

[0026] According to some embodiments of the first aspect of the present application, the process of obtaining the channel attention weight according to the first feature data includes:

[0027] Perform global average pooling on the first feature data;

[0028] Perform linear processing on the first feature data after global average pooling;

[0029] Input the first feature data after linear processing into a first activation function to obtain the channel attention weight.

[0030] According to some embodiments of the first aspect of the present application, the process of obtaining the spatial attention weight according to the second feature data includes:

[0031] Perform group normalization on the second feature data;

[0032] Perform linear processing on the second feature data after group normalization;

[0033] Input the second feature data after linear processing into a first activation function to obtain the spatial attention weight.

[0034] According to some embodiments of the first aspect of the present application, the fusion processing of the at least N-level feature maps to obtain a fused feature map includes:

[0035] Sort the N-level feature maps in descending order according to the scale of the spatial-spectral features;

[0036] Upsample the Nth-level feature map to expand the scale of the spatial-spectral features of the Nth-level feature map to be equal to the scale of the spatial-spectral features of the (N - 1)th-level feature map;

[0037] Add the upsampled Nth-level feature map to the (N - 1)th-level feature map and input the result into a second activation function to obtain the first addition result;

[0038] Upsample the Sth addition result to expand the scale of the spatial-spectral features of the Sth addition result to be equal to the scale of the spatial-spectral features of the (N - 1 - S)th-level feature map; where S = 1, 2.....N - 2;

[0039] Add the upsampled Sth addition result to the (N - 1 - S)th-level feature map and input the result into a second activation function to obtain the (S + 1)th addition result;

[0040] Until S = N - 2, take the (S + 1)th addition result as the fused feature map.

[0041] Embodiments of the second aspect of the present application provide a hyperspectral image classification method, including:

[0042] Obtain the fused feature map according to the hyperspectral image feature processing method described in the embodiments of the first aspect of the present application;

[0043] Classify the hyperspectral image according to the fused feature map.

[0044] Embodiments of the third aspect of the present application provide a hyperspectral image feature processing device, including:

[0045] A receiving unit for receiving a hyperspectral image;

[0046] A feature extraction unit for respectively inputting the spatial-spectral domain information of the hyperspectral image into N feature extraction modules to correspondingly obtain at least N-level feature maps; where N is a positive integer greater than or equal to 2, and the at least N-level feature maps have spatial-spectral features of different scales;

[0047] A fusion unit for performing fusion processing on the at least N-level feature maps to obtain a fused feature map.

[0048] The fourth aspect of the present application provides a hyperspectral image processing system, including:

[0049] At least one memory;

[0050] At least one processor;

[0051] At least one program;

[0052] The program is stored in the memory, and the processor executes at least one of the programs to implement:

[0053] The hyperspectral image feature processing method as described in the first aspect of the present application; or,

[0054] The hyperspectral image classification method as described in the second aspect of the present application.

[0055] The fifth aspect of the present application provides a computer-readable storage medium storing computer-executable instructions for executing:

[0056] The hyperspectral image feature processing method as described in the first aspect of the present application; or,

[0057] The hyperspectral image classification method as described in the second aspect of the present application.

[0058] Additional aspects and advantages of the present application will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present application. Description of the Drawings

[0059] The additional aspects and advantages of the present application will become apparent and be easily understood in conjunction with the following description of the embodiments with reference to the accompanying drawings, where:

[0060] Figure 1 It is a schematic diagram of the device architecture for executing the hyperspectral image feature processing method according to the embodiment of the present application;

[0061] Figure 2 It is a step diagram of the hyperspectral image feature processing method according to the embodiment of the present application;

[0062] Figure 3 It is a step diagram of feature extraction of the hyperspectral image feature processing method according to the embodiment of the present application;

[0063] Figure 4 It is a step diagram of feature extraction of the hyperspectral image feature processing method according to the embodiment of the present application;

[0064] Figure 5 It is an overall framework diagram of the hyperspectral image feature processing method according to the embodiment of the present application;

[0065] Figure 6 It is a step diagram of the attention update process for the hyperspectral image feature processing method of the embodiment of the present application;

[0066] Figure 7 It is a step diagram of the attention update process for the hyperspectral image feature processing method of the embodiment of the present application;

[0067] Figure 8 It is a step diagram of the attention update process for the hyperspectral image feature processing method of the embodiment of the present application;

[0068] Figure 9 It is a step diagram of the fusion for the hyperspectral image feature processing method of the embodiment of the present application;

[0069] Figure 10 It is a step diagram of the hyperspectral image classification method of the embodiment of the present application;

[0070] Figure 11 It is a block diagram of the hyperspectral image processing system of the embodiment of the present application. Detailed implementation manners

[0071] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0072] It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order from that in the flowchart. Terms such as those in the description of the specification, claims and the above drawings are used to distinguish similar objects and do not necessarily need to describe a specific order or sequence.

[0073] In the description of the present application, if the first and second are described only for the purpose of distinguishing technical features, they cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features or implicitly indicating the sequence of the indicated technical features.

[0074] In the description of the present application, unless otherwise clearly defined, words such as "set", "installed", "connected", etc. should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meanings of the above words in the present application in combination with the specific content of the technical solution.

[0075] In the related art, currently, by adopting the few-shot learning algorithm, the problem of scarce labeled samples in hyperspectral images can be alleviated to a certain extent. As a type of transfer learning, the few-shot learning algorithm only relies on a small number of labeled samples during training and can quickly generalize to new tasks with only a small number of samples with supervision information, thus solving the problem of the lack of labeled samples in hyperspectral images. However, the current few-shot learning algorithms usually use features of a single scale for prediction, ignoring features of other scales, resulting in a relatively low accuracy in subsequent hyperspectral image classification.

[0076] Based on this, the embodiments of the present application provide a hyperspectral image feature processing method, a classification method, an apparatus, a system, and a storage medium, which can obtain a fused feature map that combines spatial-spectral features of multiple different scales and has stronger discriminative ability compared with the spatial-spectral feature map of a single scale. Due to the complexity of the shooting area of hyperspectral images and the danger of field mapping, obtaining a large number of labeled samples of hyperspectral images requires a lot of manpower and materials, and the shortage of labeled samples greatly limits the application of existing deep learning algorithms to hyperspectral images.

[0077] Refer to Figure 1 , Figure 1 which is a schematic diagram of the apparatus architecture for executing the hyperspectral image feature processing method provided by the embodiments of the present application. In the Figure 1 example, the apparatus architecture includes a receiving unit, a feature extraction unit, and a fusion unit.

[0078] Among them, the receiving unit is communicatively connected to the feature extraction unit, and the fusion unit is communicatively connected to the feature extraction unit.

[0079] The receiving unit is used to connect to an external device. For example, it can be connected to an external device through Bluetooth, Wi-Fi, or other communication methods. The receiving unit is used to receive a hyperspectral image from the external device and then send the hyperspectral image to the feature extraction unit.

[0080] The feature extraction unit includes N feature extraction modules. The hyperspectral image is subjected to feature extraction through the N feature extraction modules to obtain at least N levels of feature maps. The at least N levels of feature maps have spatial-spectral features of different scales. N is a positive integer greater than or equal to 2. Then, the at least N levels of feature maps are sent to the fusion unit. The S-th feature extraction module includes S + 1 convolutional blocks and S attention blocks, where S = 1, 2.....N. For example, when N = 3, the feature extraction unit includes 3 feature extraction modules, namely the 1st feature extraction module, the 2nd feature extraction module, and the 3rd feature extraction module. The 1st feature extraction module includes 2 convolutional blocks and 1 attention block, the 2nd feature extraction module includes 3 convolutional blocks and 2 attention blocks, and the 3rd feature extraction module includes 4 convolutional blocks and 3 attention blocks. Each 1 attention block is arranged between every 2 convolutional blocks. The spatial-spectral domain information of the hyperspectral image is input into the 1st feature extraction module to obtain the 1st level of feature map, input into the 2nd feature extraction module to obtain the 2nd level of feature map, and input into the 3rd feature extraction module to obtain the 3rd level of feature map. When the spatial-spectral domain information of the hyperspectral image is input into the 3 feature extraction modules, the hyperspectral image is subjected to convolutional processing through the convolutional blocks to reduce the scale of the spatial-spectral features of the hyperspectral image. Each time convolutional processing is performed, the scale of the spatial-spectral features of the hyperspectral image is reduced by half. The hyperspectral image is subjected to attention weight update processing through the attention blocks, which can enable the obtained feature map to focus on the spatial-spectral feature components that have a greater impact on the classification result and improve the weight of the spatial-spectral feature components. Since each time convolutional processing is performed, the scale of the spatial-spectral features of the hyperspectral image is reduced by half, the scale of the spatial-spectral features of the 3rd level of feature map is half of the scale of the spatial-spectral features of the 2nd level of feature map, and the scale of the spatial-spectral features of the 2nd level of feature map is half of the scale of the spatial-spectral features of the 1st level of feature map.

[0081] The fusion unit includes an upsampling module, an addition module, and an activation module. The feature map is upsampled through the upsampling module to double the scale of the spatial-spectral features of the feature map; two feature maps with the same spatial-spectral feature scale are added through the addition module; the added feature map is subjected to non-linear processing through the activation module, which can map the feature vector of the feature map to a non-linear space, making the result of adding the feature maps more generalizable. The activation function used by the activation module is the ReLU (Rectified Linear Unit) function. For example, the process of fusing the 1st level of feature map, the 2nd level of feature map, and the 3rd level of feature map is as follows:

[0082] The 3rd level of feature map is upsampled through the upsampling module to expand the scale of the spatial-spectral features of the 3rd level of feature map to be equal to the scale of the spatial-spectral features of the 2nd level of feature map;

[0083] The third-level feature map after upsampling is added to the second-level feature map through an addition module, and nonlinear processing is performed through an activation module to obtain the first addition result;

[0084] The first addition result is upsampled through an upsampling module to expand the scale of the empty spectral features of the first addition result to be equal to the scale of the empty spectral features of the first-level feature map;

[0085] The upsampled first addition result is added to the first-level feature map through an addition module, and nonlinear processing is performed through an activation module to obtain the second addition result;

[0086] The second addition result is used as the fused feature map.

[0087] The device architecture and application scenarios described in the embodiments of the present application are for more clearly illustrating the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that with the evolution of the device architecture and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0088] Those skilled in the art can understand that Figure 1 the device architecture shown in does not constitute a limitation on the embodiments of the present application, and may include more or fewer components than shown in the figure, or combine some components, or have different component arrangements.

[0089] In Figure 1 the shown device architecture, each unit can separately call the hyperspectral image feature processing program stored therein to execute the hyperspectral image feature processing method.

[0090] Referring to Figure 2 , in a first aspect, the embodiments of the present application provide a hyperspectral image feature processing method that can be applied to Figure 1 the shown device architecture, and the hyperspectral image feature processing method includes but is not limited to step S100, step S200, and step S300.

[0091] Step S100, receive a hyperspectral image;

[0092] For example, through a hyperspectral sensor mounted on different space platforms, that is, an imaging spectrometer, in the ultraviolet, visible, near-infrared, and mid-infrared regions of the electromagnetic spectrum, the target area is simultaneously imaged in dozens to hundreds of continuous and subdivided spectral bands, thereby obtaining a hyperspectral image. The receiving unit is communicatively connected to the imaging spectrometer. After the imaging spectrometer obtains the hyperspectral image, it sends the hyperspectral image to the receiving unit so that the receiving unit can receive the hyperspectral image.

[0093] Step S200: Input the empty spectral domain information of the hyperspectral image into N feature extraction modules respectively to obtain at least N levels of feature maps; where N is a positive integer greater than or equal to 2, and the at least N levels of feature maps have empty spectral features of different scales.

[0094] In step S200, local encoding can be performed on the hyperspectral image first. For example, local encoding with a preset dimension is performed on the hyperspectral image to fuse the spatial information of the hyperspectral image. Then, the empty spectral domain information of the hyperspectral image after local encoding is input into N feature extraction modules respectively to obtain at least N levels of feature maps. The at least N levels of feature maps have empty spectral features of different scales, the sample dimensions of each level of feature map are equal, and the channel dimensions of each level of feature map are equal.

[0095] Step S300: Perform fusion processing on at least N levels of feature maps to obtain a fused feature map.

[0096] The hyperspectral image feature processing method of the embodiments of the present application can be applied to small sample learning of hyperspectral images. First, receive the hyperspectral image, then input the empty spectral domain information of the hyperspectral image into N feature extraction modules respectively to obtain at least N levels of feature maps, and then perform fusion processing on at least N levels of feature maps to obtain a fused feature map, where N is a positive integer greater than or equal to 2, and the at least N levels of feature maps have empty spectral features of different scales. Compared with a single-scale empty spectral feature map, the fused feature map that fuses multiple empty spectral features of different scales has more prominent discriminative ability, which is beneficial to improving the accuracy of subsequent classification of hyperspectral images.

[0097] In a Convolutional Neural Network (CNN), the underlying features have higher resolution and contain more fine-grained information, but their semanticity is lower and there is more noise. The high-level features have stronger semantic information, but the resolution is very low and the ability to perceive details is poor. Efficiently fusing features of multiple scales can effectively improve the performance of the model. Multi-scale feature fusion was initially widely applied in object detection and image segmentation. The purpose of feature fusion in the embodiments of the present application is to combine the empty spectral features of different scales extracted from the hyperspectral image into a feature that is more discriminative than the input features.

[0098] It can be understood that the feature extraction module includes at least one convolutional block and at least one attention block; referring to Figure 3 , step S200 may include but is not limited to step S210 and step S220.

[0099] Step S210: Input the empty spectral domain information of the hyperspectral image into the S-th feature extraction module; where S = 1, 2.....N;

[0100] Step S220: Convolve the hyperspectral image through a convolutional block, and update the hyperspectral image using attention weights through an attention block to obtain the S-level feature map.

[0101] Convolve the hyperspectral image through a convolutional block to reduce the scale of the spatial-spectral features of the hyperspectral image. Each time convolution is performed, the scale of the spatial-spectral domain information of the hyperspectral image is reduced by half. The scale of the spatial-spectral domain information is the scale of the spatial-spectral features. By updating the hyperspectral image using attention weights through an attention block, the obtained feature map can focus on the spatial-spectral feature components that have a greater impact on the classification result, and improve the weight of the spatial-spectral feature components.

[0102] It can be understood that the S-th feature extraction module includes S + 1 convolutional blocks and S attention blocks; referring to Figure 4 , step S220 may include but is not limited to step S221, step S222, step S223, and step S224.

[0103] Step S221: Perform the first convolution on the hyperspectral image through the first convolutional block to obtain the result of the first convolution; wherein, the scale of the spatial-spectral features of the result of the first convolution is half of the scale of the spatial-spectral domain information of the hyperspectral image;

[0104] Step S222: Perform the L-th attention weight update on the result of the L-th convolution through the L-th attention block to obtain the feature map after the L-th attention weight update; wherein, L = 1, 2,.....S;

[0105] Step S223: Perform the (L + 1)-th convolution on the feature map after the L-th attention weight update through the (L + 1)-th convolutional block to obtain the result of the (L + 1)-th convolution; wherein, the scale of the spatial-spectral features of the result of the (L + 1)-th convolution is half of the scale of the spatial-spectral features of the feature map after the L-th attention weight update;

[0106] Step S224: Until L = S, take the result of the (L + 1)-th convolution as the S-level feature map.

[0107] For example, when N = 3, referring to Figure 5 , Figure 5As shown in the overall framework diagram of the hyperspectral image feature processing method according to the embodiment of the present application when N = 3. Since N = 3, then S = 1, 2, 3, and there are 3 feature extraction modules, namely the first feature extraction module, the second feature extraction module, and the third feature extraction module. The first feature extraction module includes 2 convolutional blocks and 1 attention block, the second feature extraction module includes 3 convolutional blocks and 2 attention blocks, and the third feature extraction module includes 4 convolutional blocks and 3 attention blocks. Each attention block is arranged between every two convolutional blocks. The spatial-spectral domain information of the hyperspectral image is input into the first feature extraction module to obtain the first-level feature map, input into the second feature extraction module to obtain the second-level feature map, and input into the third feature extraction module to obtain the third-level feature map. For example, when S = 3, L = 1, 2, 3, the process of inputting the spatial-spectral domain of the hyperspectral image into the third feature extraction module to obtain the third-level feature map is as follows:

[0108] The first convolutional block performs the first convolution processing on the hyperspectral image to obtain the result of the first convolution processing; the scale of the spatial-spectral features of the result of the first convolution processing is half of the scale of the spatial-spectral domain information of the hyperspectral image;

[0109] The first attention block performs the first attention weight update processing on the result of the first convolution processing to obtain the feature map after the first attention weight update processing;

[0110] The second convolutional block performs the second convolution processing on the feature map after the first attention weight update processing to obtain the result of the second convolution processing; the scale of the spatial-spectral features of the result of the second convolution processing is half of the scale of the spatial-spectral features after the first attention weight update processing;

[0111] The second attention block performs the second attention weight update processing on the result of the second convolution processing to obtain the feature map after the second attention weight update processing;

[0112] The third convolutional block performs the third convolution processing on the feature map after the second attention weight update processing to obtain the result of the third convolution processing; the scale of the spatial-spectral features of the result of the third convolution processing is half of the scale of the spatial-spectral features of the feature map after the second attention weight update processing;

[0113] The third attention block performs the third attention weight update processing on the result of the third convolution processing to obtain the feature map after the third attention weight update processing;

[0114] The feature map after the 3rd attention weight update process is subjected to the 4th convolution process through the 4th convolution block to obtain the result of the 4th convolution process; the scale of the empty spectrum feature of the result of the 4th convolution process is half of the scale of the empty spectrum feature of the feature map after the 3rd attention weight update process;

[0115] The result of the 4th convolution process is used as the feature map of the 3rd level.

[0116] It can be understood that the convolution block includes a convolution layer (conv), a normalization layer (batchnormlize), an activation function layer, and a max pooling layer (maxpool2d). The convolution layer can be an n-d convolution layer, representing an n-dimensional convolution layer, and n can take a value of 1 or more. For example, if n takes a value of 1, the calculation formula in the 1-d convolution layer is:

[0117]

[0118] In the calculation formula of the 1-d convolution layer, v is the feature map output by the convolution, B and M are the sizes of the convolution kernel along the empty spectrum feature dimension, r is the bias matrix, k is the convolution kernel matrix, b and m are the indices of k respectively, z is the index of the feature map, i is the index of the feature map, j is the index of the feature map, f uses ReLU as the activation function, and the function form of ReLU is as follows:

[0119] f(x) = max(0, x)

[0120] x is the input of ReLU.

[0121] It can be understood that referring to Figure 6 , step S222 may include but is not limited to step S500, step S510, step S520, step S530, and step S540.

[0122] Step S500, dividing the result of the Lth convolution process into several groups of feature data in the channel dimension;

[0123] Step S510, dividing each group of feature data into first feature data and second feature data;

[0124] Step S520, obtaining the channel attention weight according to the first feature data;

[0125] Step S530, obtaining the spatial attention weight according to the second feature data;

[0126] Step S540, obtaining the feature map after the Lth attention weight update process according to the channel attention weight and the spatial attention weight.

[0127] For example, the result of the Lth convolution process is X (B，C，H), where B is the sample dimension, C is the channel dimension, and H is the scale of the empty spectral feature. The results of L convolutional processes are divided into g groups of feature data in the channel dimension. g is a positive integer greater than or equal to 2. Then the number of channels in each group of feature data is C / g. The result of the L-th convolutional process can be expressed as:

[0128] X (B,C,H) ={X 1(B,C / g,H) ,X 2(B,C / g,H) ......,X g(B,C / g,H)}

[0129] Each group of feature data is divided into first feature data and second feature data. The channel attention weight of the input feature map is obtained according to the first feature data, and the spatial attention weight of the input feature map is obtained according to the second feature data. Then, according to the channel attention weight and the spatial attention weight, the channel attention weights and spatial attention weights of all groups are directly concatenated, and the order of the channel attention weights is randomly swapped in the channel dimension to obtain the feature map after the L-th attention weight update process. In this way, the feature map after the L-th attention weight update process can improve the weight of the spatial-spectral feature components of the result of the L-th convolutional process, and because the order of the channel attention weights is randomly swapped in the channel dimension, the feature map after the L-th attention weight update process has a certain learning ability.

[0130] It can be understood that referring to Figure 7 , step S520 may include but is not limited to step S521, step S522, and step S523.

[0131] Step S521, perform global average pooling on the first feature data;

[0132] Step S522, perform linear processing on the first feature data after global average pooling;

[0133] Step S523, input the first feature data after linear processing into the first activation function to obtain the channel attention weight. The first activation function is the sigmoid function.

[0134] It can be understood that referring to Figure 8 , step S530 may include but is not limited to step S531, step S532, and step S533.

[0135] Step S531, perform group normalization (GroupNorm) on the second feature data;

[0136] The specific calculation formula of group normalization is:

[0137]

[0138] Among them, X represents the input data, Y represents the output data, E represents the averaging operator, Var represents the standard deviation operator, γ and β are learnable mapping transformation parameters, and ε is used to prevent the denominator from being zero.

[0139] Step S532: Linearly process the second feature data after group normalization.

[0140] Step S533: Input the linearly processed second feature data into the first activation function to obtain the spatial attention weight.

[0141] It can be understood that referring to Figure 9 , step S300 may include but is not limited to step S310, step S320, step S330, step S340, step S350, and step S360.

[0142] Step S310: Sort the N-level feature maps in descending order according to the scale of the spatial spectral features.

[0143] Step S320: Upsample the N-level feature map to expand the scale of the spatial spectral features of the N-level feature map to be equal to the scale of the spatial spectral features of the N-1-level feature map.

[0144] Step S330: Add the upsampled N-level feature map to the N-1-level feature map and input the result into the second activation function to obtain the first addition result.

[0145] In step S330, the second activation function is the ReLU (Rectified Linear Unit) function. By performing non-linear processing on the feature map through the second activation function, the feature vectors of the feature map can be mapped into a non-linear space, making the first addition result more generalizable.

[0146] Step S340: Upsample the S-th addition result to expand the scale of the spatial spectral features of the S-th addition result to be equal to the scale of the spatial spectral features of the N-1-S-level feature map; where S = 1, 2.....N-2.

[0147] Step S350: Add the upsampled S-th addition result to the N-1-S-level feature map and input the result into the second activation function to obtain the S+1-th addition result.

[0148] Step S360: Until S = N-2, use the S+1-th addition result as the fused feature map.

[0149] For example, referring to Figure 5, N = 3. The spatial-spectral domain information of the hyperspectral image is input into the first feature extraction module to obtain the first-level feature map, input into the second feature extraction module to obtain the second-level feature map, and input into the third feature extraction module to obtain the third-level feature map. The scales of the spatial-spectral features of the first-level feature map, the second-level feature map, and the third-level feature map increase gradually. The process of fusing the first-level feature map, the second-level feature map, and the third-level feature map to obtain the fused feature map is as follows:

[0150] Upsample the third-level feature map to expand the scale of the spatial-spectral features of the third-level feature map to be equal to the scale of the spatial-spectral features of the second-level feature map;

[0151] Add the upsampled third-level feature map and the second-level feature map and input them into the second activation function to obtain the first addition result;

[0152] Upsample the first addition result to expand the scale of the spatial-spectral features of the first addition result to be equal to the scale of the spatial-spectral features of the first-level feature map;

[0153] Add the upsampled first addition result and the first-level feature map and input them into the second activation function to obtain the second addition result;

[0154] Take the second addition result as the fused feature map.

[0155] It can be understood that for the hyperspectral image feature processing method of the embodiments of the present application, after obtaining the fused feature map, the fused feature map is converted into a one-dimensional tensor in the spectral dimension and the channel dimension to obtain a fused feature vector. For example, the scale of the fused feature is 200×32×12, where 200 is the sample dimension, 32 is the channel dimension, and 12 is the scale of the spatial-spectral features. After being converted into a one-dimensional tensor in the spectral dimension and the channel dimension, the scale of the fused feature vector is 200×384. The fused feature vector can be used to calculate various prototypes of the hyperspectral image. The various prototypes calculated through the fused feature vector have higher accuracy when performing classification and recognition.

[0156] Next, with reference to Figures 1 to 9 , a specific embodiment is used to describe in detail the hyperspectral image feature processing method applying the embodiments of the first aspect of the present application. It should be understood that the following description is only an exemplary illustration and not a specific limitation of the present application.

[0157] Figure 5As shown in the overall framework diagram of the hyperspectral image feature processing method according to the embodiment of the present application when N = 3. Since N = 3, then S = 1, 2, 3, so there are 3 feature extraction modules, namely the first feature extraction module, the second feature extraction module, and the third feature extraction module. The first feature extraction module includes 2 convolutional blocks and 1 attention block, the second feature extraction module includes 3 convolutional blocks and 2 attention blocks, and the third feature extraction module includes 4 convolutional blocks and 3 attention blocks. Each attention block is arranged between every 2 convolutional blocks.

[0158] For the hyperspectral image feature processing method according to the embodiment of the present application, first, the hyperspectral image is locally encoded to fuse the spatial information of the hyperspectral image. Then, the spatial-spectral domain information of the locally encoded hyperspectral image is respectively input into the first feature extraction module, the second feature extraction module, and the third feature extraction module to obtain the first-level feature map, the second-level feature map, and the third-level feature map. The convolutional blocks are used to perform convolutional processing on the hyperspectral image to reduce the scale of the spatial-spectral domain information of the hyperspectral image. Each time convolutional processing is performed, the scale of the spatial-spectral domain information of the hyperspectral image is reduced by half. The attention blocks are used to update the hyperspectral image with attention weights, which can enable the obtained feature map to focus on the spatial-spectral feature components that have a greater impact on the classification result and improve the weight of the spatial-spectral feature components.

[0159] Each time the hyperspectral image undergoes convolutional processing by a convolutional block, the scale of the spatial-spectral features is reduced by half. Suppose the size of the hyperspectral image is (B, C, H). B is the sample dimension, C is the channel dimension, and H is the scale of the spatial-spectral features. Then the size of the first-level feature map is (B, C, H / 4), the size of the second-level feature map is (B, C, H / 8), and the size of the third-level feature map is (B, C, H / 16); for example, the process of inputting the spatial-spectral domain information of the hyperspectral image into the third feature extraction module to obtain the third-level feature map is as follows:

[0160] The first convolutional block performs the first convolutional processing on the hyperspectral image to obtain the result of the first convolutional processing; the scale of the spatial-spectral features of the result of the first convolutional processing is half of the scale of the spatial-spectral domain information of the hyperspectral image, and the size of the result of the first convolutional processing is (B, C, H / 2);

[0161] The first attention block performs the first attention weight update processing on the result of the first convolutional processing to obtain the feature map after the first attention weight update processing;

[0162] The feature map after the first attention weight update process is subjected to a second convolution process through the second convolution block to obtain the result of the second convolution process; the scale of the empty spectral feature of the result of the second convolution process is half of the scale of the empty spectral feature of the feature map after the first attention weight update process, and the size of the result of the second convolution process is (B, C, H / 4);

[0163] The result of the second convolution process is subjected to a second attention weight update process through the second attention block to obtain the feature map after the second attention weight update process;

[0164] The feature map after the second attention weight update process is subjected to a third convolution process through the third convolution block to obtain the result of the third convolution process; the scale of the empty spectral feature of the result of the third convolution process is half of the scale of the empty spectral feature of the feature map after the second attention weight update process, and the size of the result of the third convolution process is (B, C, H / 8);

[0165] The result of the third convolution process is subjected to a third attention weight update process through the third attention block to obtain the feature map after the third attention weight update process;

[0166] The feature map after the third attention weight update process is subjected to a fourth convolution process through the fourth convolution block to obtain the result of the fourth convolution process; the scale of the empty spectral feature of the result of the fourth convolution process is half of the scale of the empty spectral feature of the feature map after the third attention weight update process, and the size of the result of the fourth convolution process is (B, C, H / 16);

[0167] The result of the fourth convolution process is used as the third-level feature map.

[0168] After obtaining the first-level feature map, the second-level feature map, and the third-level feature map, the first-level feature map, the second-level feature map, and the third-level feature map are fused, and the process is as follows:

[0169] The third-level feature map is upsampled to expand the scale of the empty spectral feature of the third-level feature map to be equal to the scale of the empty spectral feature of the second-level feature map;

[0170] The upsampled third-level feature map and the second-level feature map are added and input to the second activation function to obtain the first addition result;

[0171] The first addition result is upsampled to expand the scale of the empty spectral feature of the first addition result to be equal to the scale of the empty spectral feature of the first-level feature map;

[0172] The upsampled first addition result and the first-level feature map are added and input to the second activation function to obtain the second addition result;

[0173] Use the result of the second addition as the fused feature map.

[0174] After obtaining the fused feature map, convert the fused feature map into a one-dimensional tensor in the spectral dimension and the channel dimension to obtain a fused feature vector. For example, the scale of the fused feature is 200×32×12, where 200 is the sample dimension, 32 is the channel dimension, and 12 is the spatial-spectral feature dimension. After converting into a one-dimensional tensor in the spatial-spectral feature dimension and the channel dimension, the scale of the fused feature vector is 200×384. The fused feature vector can be used to calculate various prototypes of the hyperspectral image. The various prototypes calculated through the fused feature vector have higher accuracy when performing classification and recognition.

[0175] In a second aspect, referring to Figure 10 , embodiments of the present application provide a hyperspectral image classification method, including but not limited to step S910 and step S920.

[0176] Step S910: Obtain a fused feature map according to the hyperspectral image feature processing method of the first aspect embodiment of the present application;

[0177] Step S920: Classify the hyperspectral image according to the fused feature map.

[0178] The hyperspectral image classification method of the embodiments of the present application first obtains a fused feature map through the hyperspectral image feature processing method of the first aspect embodiment, and then according to the fused feature map, converts the fused feature map into a one-dimensional tensor in the spatial-spectral feature dimension and the channel dimension to obtain a fused feature vector. For example, the scale of the fused feature is 200×32×12, where 200 is the sample dimension, 32 is the channel dimension, and 12 is the spatial-spectral feature dimension. The spatial-spectral feature dimension is the scale of the spatial-spectral feature. After converting into a one-dimensional tensor in the spatial-spectral feature dimension and the channel dimension, the scale of the fused feature vector is 200×384. The fused feature vector can be used to calculate various prototypes of the hyperspectral image. When performing classification and recognition, the hyperspectral image is compared with the prototype image to obtain a classification result. In the related art, usually a single-scale spatial-spectral feature map is used for classification, while the classification method of the embodiments of the present application uses a fused feature map that fuses multiple different-scale spatial-spectral features, and the discrimination ability of the fused feature map is more prominent, making the accuracy of the hyperspectral image classification method of the embodiments of the present application higher.

[0179] It can be understood that, referring to Table 1, the classification method of the embodiment of the present application is tested on the PU (Pavia University) dataset and compared with related methods in the field. The accuracy of the classification method is tested by two test methods: 3-shot classification and 5-shot classification. 3-shot classification means taking 3 samples of each class as support samples, and the remaining samples as query samples; 5-shot classification means taking 5 samples of each class as support samples, and the remaining samples as query samples. It can be seen from the test results shown in Table 1 that the accuracy of the classification method of the embodiment of the present application is higher than that of related methods in the field. Among them, 1D-CNN represents a one-dimensional convolutional neural network, 2D-CNN represents a two-dimensional convolutional neural network, and 3D-CNN represents a three-dimensional convolutional neural network.

[0180] Table 1

[0181]

[0182] In a third aspect, an embodiment of the present application provides a hyperspectral image feature processing device, including:

[0183] A receiving unit for receiving a hyperspectral image;

[0184] A feature extraction unit for respectively inputting the spatial-spectral domain information of the hyperspectral image into N feature extraction modules to correspondingly obtain at least N levels of feature maps; where N is a positive integer greater than or equal to 2, and the at least N levels of feature maps have spatial-spectral features of different scales;

[0185] A fusion unit for performing fusion processing on at least N levels of feature maps to obtain a fused feature map.

[0186] The hyperspectral image feature processing device of the embodiment of the present application receives a hyperspectral image through the receiving unit, then respectively inputs the spatial-spectral domain features of the hyperspectral image into N feature extraction modules through the feature extraction unit to correspondingly obtain at least N levels of feature maps, and then performs fusion processing on at least N levels of feature maps to obtain a fused feature map. Since the at least N levels of feature maps have spatial-spectral features of different scales, compared with a single-scale spatial-spectral feature map, the discriminative ability of the fused feature map that fuses multiple spatial-spectral features of different scales is more prominent, which is beneficial to improving the accuracy of subsequent classification of hyperspectral images.

[0187] It should be noted that the hyperspectral image feature processing device in the above-mentioned embodiments is based on the same inventive concept as the hyperspectral image feature processing method in the above-mentioned embodiments. Therefore, the corresponding content of the hyperspectral image feature processing method in the above-mentioned embodiments is equally applicable to the hyperspectral image feature processing device in the above-mentioned embodiments, and has the same implementation principle and technical effects. To avoid redundant description, it will not be described in detail here.

[0188] Fourthly, referring to Figure 11 , an embodiment of the present application provides a hyperspectral image processing system, including:

[0189] At least one memory 200;

[0190] At least one processor 100;

[0191] At least one program;

[0192] The program is stored in the memory 200, and the processor 100 executes at least one program to implement:

[0193] The hyperspectral image feature processing method according to the first aspect embodiment of the present application: or,

[0194] The hyperspectral image classification method according to the second aspect embodiment of the present application.

[0195] The processor 100 and the memory 200 can be connected through a bus or other means.

[0196] The memory 200, as a non-transitory readable storage medium, can be used to store non-transitory software instructions and non-transitory executable instructions. In addition, the memory 200 may include a high-speed random access memory 200, and may also include a non-transitory memory 200, such as at least one disk memory 200, a flash memory device, or other non-transitory solid-state memory 200. It can be understood that the memory 200 may optionally include a memory 200 remotely provided with respect to the processor 100, and these remote memories 200 can be connected to the processor 100 through a network. Examples of the above network include but are not limited to the Internet, enterprise intranet, local area network, mobile communication network and their combinations.

[0197] The processor 100 realizes various functional applications and data processing by running non-transitory software instructions, instructions and signals stored in the memory 200, that is, realizes the hyperspectral image feature processing method according to the first aspect embodiment or the hyperspectral image classification method according to the second aspect embodiment.

[0198] The non-transitory software instructions required to implement the hyperspectral image feature processing method of the first aspect embodiment or the hyperspectral image classification method of the second aspect embodiment, and the instructions are stored in the memory 200. When executed by the processor 100, the hyperspectral image feature processing method of the first aspect embodiment or the hyperspectral image classification method of the second aspect embodiment of the present application is executed. For example, execute the Figure 2 method steps S100 to S300 in Figure 3 method steps S210 to S220 in Figure 4 method steps S221 to S224 in Figure 6 method steps S500 to S540 in Figure 7 method steps S521 to S523 in Figure 8 method steps S531 to S533 in Figure 9 method steps S310 to S360 in Figure 10 method steps S910 to S920 in

[0199] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0200] In addition, the embodiments of the present application also provide a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are executed by a processor or a controller, for example, executed by a processor of the hyperspectral image processing system in the above embodiments, the above processor can execute the hyperspectral image feature processing method of the first aspect embodiment of the present application or the hyperspectral image classification method of the second aspect embodiment of the present application. For example, execute the Figure 2 method steps S100 to S300 in Figure 3 method steps S210 to S220 in Figure 4 method steps S221 to S224 in Figure 6 method steps S500 to S540 in Figure 7 method steps S521 to S523 in Figure 8 method steps S531 to S533 in Figure 9 method steps S310 to S360 in Figure 10 method steps S910 to S920 in

[0201] Through the description of the above embodiments, those of ordinary skill in the art can understand that all or some of the steps and systems disclosed above can be implemented as software, firmware, hardware, and their appropriate combinations. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or can be implemented as hardware, or can be implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a readable medium, which can include a computer storage medium (or non-transitory medium) and a communication medium (or transitory medium). As is well known to those of ordinary skill in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable signals, data structures, instruction modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those of ordinary skill in the art that a communication medium typically contains computer-readable signals, data structures, instruction modules, or other data in a modulated data signal such as a carrier wave or other transmission mechanism, and can include any information delivery medium.

[0202] The embodiments of the present application have been described in detail above in conjunction with the accompanying drawings. However, the present application is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the purpose of the present application.

Claims

1. A method for processing hyperspectral image features, characterized in that, comprising: Receiving a hyperspectral image; Inputting the spatial-spectral domain information of the hyperspectral image into N feature extraction modules respectively, and correspondingly obtaining at least N levels of feature maps; wherein, N is a positive integer greater than or equal to 2, and the at least N levels of feature maps have spatial-spectral features of different scales; Performing fusion processing on the at least N levels of feature maps to obtain a fused feature map; The feature extraction module includes at least one convolutional block and at least one attention block; The step of inputting the spatial-spectral domain information of the hyperspectral image into N feature extraction modules respectively and correspondingly obtaining at least N levels of feature maps includes: Inputting the spatial-spectral domain information of the hyperspectral image into the S-th feature extraction module; wherein, S = 1, 2.....N; Performing convolutional processing on the hyperspectral image through the convolutional block, and performing update processing on the hyperspectral image using attention weights through the attention block to obtain the S-th level of feature map; The S-th feature extraction module includes S + 1 convolutional blocks and S attention blocks; The step of performing convolutional processing on the hyperspectral image through the convolutional block and performing update processing on the hyperspectral image using attention weights through the attention block to obtain the S-th level of feature map includes: Performing the first convolutional processing on the hyperspectral image through the first convolutional block to obtain the result of the first convolutional processing; wherein, the scale of the spatial-spectral features of the result of the first convolutional processing is half of the scale of the spatial-spectral domain information of the hyperspectral image; Performing the L-th attention weight update processing on the result of the L-th convolutional processing through the L-th attention block to obtain the feature map after the L-th attention weight update processing; wherein, L = 1, 2,.....S; Performing the (L + 1)-th convolutional processing on the feature map after the L-th attention weight update through the (L + 1)-th convolutional block to obtain the result of the (L + 1)-th convolutional processing; wherein, the scale of the spatial-spectral features of the result of the (L + 1)-th convolutional processing is half of the scale of the spatial-spectral features of the feature map after the L-th attention weight update processing; Until L = S, taking the result of the (L + 1)-th convolutional processing as the S-th level of feature map.

2. The method for processing hyperspectral image features according to claim 1, characterized in that, The step of performing the L-th attention weight update processing on the result of the L-th convolutional processing through the L-th attention block to obtain the feature map after the L-th attention weight update processing includes: Dividing the result of the L-th convolutional processing into several groups of feature data in the channel dimension; Dividing each group of the feature data into first feature data and second feature data; Obtaining channel attention weights according to the first feature data; Obtaining spatial attention weights according to the second feature data; Obtaining the feature map after the L-th attention weight update processing according to the channel attention weights and the spatial attention weights.

3. The method for processing hyperspectral image features according to claim 2, characterized in that, The step of obtaining channel attention weights according to the first feature data includes: Performing global average pooling on the first feature data; Perform linear processing on the first feature data after global average pooling; Input the first feature data after linear processing into a first activation function to obtain the channel attention weight.

4. The hyperspectral image feature processing method according to claim 2, wherein, obtaining the spatial attention weight according to the second feature data includes: Performing group normalization on the second feature data; Performing linear processing on the second feature data after group normalization; Input the second feature data after linear processing into a first activation function to obtain the spatial attention weight.

5. The hyperspectral image feature processing method according to claim 1, wherein, performing fusion processing on the at least N-level feature maps to obtain a fused feature map includes: Sorting the N-level feature maps in descending order according to the scale of the spatial-spectral features; Performing upsampling on the Nth-level feature map to expand the scale of the spatial-spectral features of the Nth-level feature map to be equal to the scale of the spatial-spectral features of the (N - 1)th-level feature map; Adding the upsampled Nth-level feature map to the (N - 1)th-level feature map and inputting the result into a second activation function to obtain the first addition result; Performing upsampling on the Sth addition result to expand the scale of the spatial-spectral features of the Sth addition result to be equal to the scale of the spatial-spectral features of the (N - 1 - S)th-level feature map; where S = 1, 2.....N - 2; Adding the upsampled Sth addition result to the (N - 1 - S)th-level feature map and inputting the result into a second activation function to obtain the (S + 1)th addition result; Until S = N - 2, taking the (S + 1)th addition result as the fused feature map.

6. A hyperspectral image classification method, wherein, includes: Obtaining the fused feature map according to the hyperspectral image feature processing method according to any one of claims 1 to 5; Classifying the hyperspectral image according to the fused feature map.

7. A hyperspectral image feature processing device, wherein, includes: A receiving unit for receiving a hyperspectral image; A feature extraction unit for respectively inputting the spatial-spectral domain information of the hyperspectral image into N feature extraction modules to correspondingly obtain at least N-level feature maps; where N is a positive integer greater than or equal to 2, and the at least N-level feature maps have spatial-spectral features of different scales; A fusion unit for performing fusion processing on the at least N-level feature maps to obtain a fused feature map; The feature extraction module includes at least one convolutional block and at least one attention block; The step of respectively inputting the spatial-spectral domain information of the hyperspectral image into N feature extraction modules to correspondingly obtain at least N-level feature maps includes: Inputting the spatial-spectral domain information of the hyperspectral image into the Sth feature extraction module; where S = 1, 2.....N; Performing convolutional processing on the hyperspectral image through the convolutional block and performing update processing on the hyperspectral image using the attention weight through the attention block to obtain the Sth-level feature map; The S-th feature extraction module includes S + 1 convolutional blocks and S attention blocks; Performing convolutional processing on the hyperspectral image through the convolutional blocks, and performing update processing on the hyperspectral image using attention weights through the attention blocks to obtain the S-th level feature map, including: Performing the first convolutional processing on the hyperspectral image through the first convolutional block to obtain the result of the first convolutional processing; wherein, the scale of the spatial-spectral features of the result of the first convolutional processing is half of the scale of the spatial-spectral domain information of the hyperspectral image; Performing the L-th attention weight update processing on the result of the L-th convolutional processing through the L-th attention block to obtain the feature map after the L-th attention weight update processing; wherein, L = 1, 2,.....S; Performing the (L + 1)-th convolutional processing on the feature map after the L-th attention weight update through the (L + 1)-th convolutional block to obtain the result of the (L + 1)-th convolutional processing; wherein, the scale of the spatial-spectral features of the result of the (L + 1)-th convolutional processing is half of the scale of the spatial-spectral features of the feature map after the L-th attention weight update processing; Until L = S, taking the result of the (L + 1)-th convolutional processing as the S-th level feature map.

8. A hyperspectral image processing system Characterized in that It includes: At least one memory; At least one processor; At least one program; The program is stored in the memory, and the processor executes at least one of the programs to implement: The hyperspectral image feature processing method according to any one of claims 1 to 5; or, The hyperspectral image classification method according to claim 6.

9. A computer-readable storage medium Characterized in that It stores computer-executable instructions, and the computer-executable instructions are used to execute: The hyperspectral image feature processing method according to any one of claims 1 to 5; or, The hyperspectral image classification method according to claim 6.

Images