A hyperspectral remote sensing image classification method and device

Abstract

The application discloses a hyperspectral remote sensing image classification method and device, and relates to the image recognition technical field. The method comprises the following steps: acquiring an image to be classified; inputting the image to be classified into a target multi-feature fusion reservation network model to obtain a hyperspectral remote sensing image classification result; wherein the target multi-feature fusion reservation network model is obtained through the following steps: acquiring an initial hyperspectral remote sensing image; preprocessing the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image; randomly sampling the target hyperspectral remote sensing image according to a preset size to obtain model input images; constructing an initial multi-feature fusion reservation network model; inputting the model input images into the initial multi-feature fusion reservation network model to train the initial multi-feature fusion reservation network model and obtain the target multi-feature fusion reservation network model. The application realizes the classification of the hyperspectral remote sensing image, improves the efficiency and accuracy, and can be widely applied in the image recognition technical field.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and in particular to a hyperspectral remote sensing image classification method and device. BACKGROUND

[0002] Hyperspectral remote sensing images record hundreds of bands of each pixel point on the electromagnetic spectrum. Different ground objects have differences in the emission, reflection and absorption of electromagnetic waves, so that hyperspectral images can identify and detect ground objects at a fine granularity level. Different ground objects have differences in the emission, reflection and absorption of electromagnetic waves, so that hyperspectral images can identify and detect ground objects at a fine granularity level. Traditional classification uses various machine learning methods, but usually relies on manual feature extraction, which is time-consuming and inefficient. At the same time, the spatial and spectral characteristics of the ground objects are ignored, and the classification accuracy is low.

[0003] In summary, the technical problems in the related art need to be improved. SUMMARY

[0004] The present application provides a hyperspectral remote sensing image classification method and device, which effectively improves the efficiency and accuracy.

[0005] In one aspect, the present application provides a hyperspectral remote sensing image classification method, comprising the following steps:

[0006] Obtaining a to-be-classified image;

[0007] Inputting the to-be-classified image into a target multi-feature fusion reservation network model to obtain a hyperspectral remote sensing image classification result;

[0008] Wherein, the target multi-feature fusion reservation network model is obtained by the following steps:

[0009] Obtaining an initial hyperspectral remote sensing image;

[0010] Pretreating the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image;

[0011] Randomly sampling the target hyperspectral remote sensing image according to a preset size to obtain a model input image;

[0012] Constructing an initial multi-feature fusion reservation network model;

[0013] Inputting a plurality of model input images into the initial multi-feature fusion reservation network model to enable the initial multi-feature fusion reservation network model to be trained to obtain the target multi-feature fusion reservation network model.

[0014] In some embodiments, the pretreating the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image comprises:

[0015] performing principal component analysis on the initial hyperspectral remote sensing image to obtain a first image, a channel number of the first image being less than a channel number of the initial hyperspectral remote sensing image;

[0016] filling edges of the first image by using a zero-marked pixel filling method according to a filling width to obtain the target hyperspectral remote sensing image.

[0017] In some embodiments, the construction process of the initial multi-feature fusion reservation network model includes:

[0018] constructing a first multi-feature fusion network, the first multi-feature fusion network being configured to extract significant features from the model input image to obtain a first multi-feature fusion network output feature map;

[0019] constructing a second multi-feature fusion network after the first multi-feature fusion network, the second multi-feature fusion network being configured to extract subtle features from the first multi-feature fusion network output feature map to obtain a second multi-feature fusion network output feature map;

[0020] constructing a reservation encoding network after the second multi-feature fusion network, the reservation encoding network being configured to extract spatial-spectral feature internal correlation from the second multi-feature fusion network output feature map;

[0021] constructing a global pooling layer after the reservation encoding network;

[0022] constructing a linear layer after the global pooling layer.

[0023] In some embodiments, the construction process of the first multi-feature fusion network includes:

[0024] constructing a first convolutional layer, a convolution kernel size of the first convolutional layer being 3x3, the first convolutional layer being configured to perform convolution on the model input image to obtain a first feature map;

[0025] constructing a first channel attention layer after the first convolutional layer, the first channel attention layer being configured to reorder the first feature map according to color depth to obtain a second feature map;

[0026] constructing a first partial convolution processing layer after the first channel attention layer, the first partial convolution processing layer being configured to perform partial convolution processing on a dark gray feature map in the second feature map according to a preset channel number to obtain a third feature map;

[0027] After the first part of the convolution processing layer, a second convolution layer is constructed, the convolution kernel size of the second convolution layer is 1x1, and the second convolution layer is used for convolution on the third feature map to obtain a fourth feature map;

[0028] After the second convolution layer, a first batch normalization layer is constructed, and the first batch normalization layer is used for batch normalization on the fourth feature map according to a preset activation function to obtain the first multi-feature fusion network output feature map.

[0029] In some embodiments, the construction process of the second multi-feature fusion network comprises:

[0030] A third convolution layer is constructed, the convolution kernel size of the third convolution layer is 3x3, and the third convolution layer is used for convolution on the first multi-feature fusion network output feature map to obtain a fifth feature map;

[0031] After the third convolution layer, a second channel attention layer is constructed, and the second channel attention layer is used for reordering the fifth feature map according to color depth to obtain a sixth feature map;

[0032] After the second channel attention layer, a second part of the convolution processing layer is constructed, and the second part of the convolution processing layer is used for partial convolution processing on a light gray feature map in the sixth feature map according to a preset channel number to obtain a seventh feature map;

[0033] After the second part of the convolution processing layer, a fourth convolution layer is constructed, the convolution kernel size of the fourth convolution layer is 1x1, and the fourth convolution layer is used for convolution on the seventh feature map to obtain an eighth feature map;

[0034] After the fourth convolution layer, a second batch normalization layer is constructed, and the second batch normalization layer is used for batch normalization on the eighth feature map according to a preset activation function to obtain the second multi-feature fusion network output feature map.

[0035] In some embodiments, the construction process of the reservation encoding network comprises:

[0036] A feature flattening layer is constructed, and the feature flattening layer is used for flattening the second multi-feature fusion network output feature map to obtain a feature sequence;

[0037] After the feature flattening layer, a simple reservation layer is constructed, and the simple reservation layer is used for feature extraction processing on the feature sequence to obtain a simple reservation layer output feature.

[0038] After the simple reservation layer, a feedforward network layer is constructed.

[0039] In some embodiments, the partial convolution processing of the dark gray feature maps in the second feature map according to the preset number of channels comprises:

[0040] dividing the second feature map according to the preset number of channels to obtain a first divided feature map and a second divided feature map, the first divided feature map having the preset number of channels, and the number of dark gray feature maps in the first divided feature map being greater than the number of dark gray feature maps in the second divided feature map;

[0041] performing convolution on the first divided feature map to obtain a to-be-stitched feature map;

[0042] stitching the to-be-stitched feature map and the second divided feature map to obtain the third feature map.

[0043] In some embodiments, the feature extraction processing of the feature sequence to obtain simple reserved layer output features comprises:

[0044] performing matrix conversion on the feature sequence to obtain a feature matrix;

[0045] calculating a query matrix according to the feature matrix and a query learnable matrix;

[0046] calculating a key matrix according to the feature matrix and a key learnable matrix;

[0047] calculating a value matrix according to the feature matrix and a value learnable matrix;

[0048] calculating a combined matrix of the causal mask and the exponential decay along the relative distance according to a scalar constant;

[0049] calculating the simple reserved layer output features according to the query matrix, the key matrix, the value matrix, and the combined matrix of the causal mask and the exponential decay along the relative distance.

[0050] In another aspect, an embodiment of the present application provides a hyperspectral remote sensing image classification device, comprising:

[0051] A first module is configured to obtain a to-be-classified image.

[0052] A second module is configured to input the to-be-classified image into a target multi-feature fusion reserved network model to obtain a hyperspectral remote sensing image classification result.

[0053] The target multi-feature fusion reserved network model is obtained by the following steps:

[0054] An initial hyperspectral remote sensing image is obtained.

[0055] Pretreat the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image;

[0056] Randomly sample the target hyperspectral remote sensing image according to a preset size to obtain a model input image;

[0057] An initial multi-feature fusion reservation network model is constructed.

[0058] The plurality of model input images are input into the initial multi-feature fusion reservation network model, so that the initial multi-feature fusion reservation network model is trained to obtain the target multi-feature fusion reservation network model.

[0059] In another aspect, an embodiment of the present application provides a computer device, comprising:

[0060] At least one processor;

[0061] At least one memory for storing at least one program;

[0062] When the at least one program is executed by the at least one processor, the at least one processor implements the method.

[0063] The present application has the following beneficial effects:

[0064] The embodiment of the present application first acquires an image to be classified, and then inputs the image to be classified into a target multi-feature fusion reservation network model to obtain a hyperspectral remote sensing image classification result, thereby realizing hyperspectral remote sensing image classification and improving efficiency and accuracy. The target multi-feature fusion reservation network model is constructed by first acquiring an initial hyperspectral remote sensing image, pretreating the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image, randomly sampling the target hyperspectral remote sensing image according to a preset size to obtain a model input image, constructing an initial multi-feature fusion reservation network model, and finally inputting a plurality of model input images into the initial multi-feature fusion reservation network model to train the initial multi-feature fusion reservation network model, thereby improving model precision through a multi-feature fusion network and a reservation coding network.

[0065] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art upon examination of the following or can be learned by practice of the present application. The objects and other advantages of the present application can be achieved and attained by the structures particularly pointed out in the description and the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0067] Figure 1 A flow chart of a hyperspectral remote sensing image classification method according to an embodiment of the present application;

[0068] Figure 2 A flow chart of obtaining a target multi-feature fusion reservation network model according to an embodiment of the present application;

[0069] Figure 3 A schematic diagram of a center point pixel according to an embodiment of the present application;

[0070] Figure 4 A schematic diagram of edge pixel filling according to an embodiment of the present application;

[0071] Figure 5 A schematic diagram of a multi-feature fusion reservation network model structure according to an embodiment of the present application;

[0072] Figure 6 A schematic diagram of a partial convolution processing process according to an embodiment of the present application;

[0073] Figure 7 A schematic diagram of a model classification precision comparison result according to an embodiment of the present application;

[0074] Figure 8 A three-channel RGB image according to an embodiment of the present application;

[0075] Figure 9 A real ground object label image according to an embodiment of the present application;

[0076] Figure 10 A HybridSn classification result image according to an embodiment of the present application;

[0077] Figure 11 A PiT classification result image according to an embodiment of the present application;

[0078] Figure 12 A HiT classification result image according to an embodiment of the present application;

[0079] Figure 13 A classification result image according to an embodiment of the present application;

[0080] Figure 14 A schematic diagram of a whole process of training a target multi-feature fusion reservation network model according to an embodiment of the present application;

[0081] Figure 15 Fig. 1 is a structural schematic diagram of a hyperspectral remote sensing image classification device according to an embodiment of the present application;

[0082] Figure 16 Fig. 2 is a hardware structural schematic diagram of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0083] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the accompanying drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all the implementations consistent with the embodiments of the present application, and they are only examples of devices and methods consistent with some aspects of the embodiments of the present application as detailed in the appended claims.

[0084] It can be understood that the terms "first", "second", and the like used in the present application can be used herein to describe various concepts, but unless specifically stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of the present application, the first information can also be referred to as the second information, and similarly, the second information can also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "when" or "when" or "in response to determining".

[0085] The terms "at least one", "multiple", "each", "any" and the like used in the present application include one, two or more than two, multiple includes two or more than two, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0086] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0087] Before the embodiments of the present application are described in detail, first, some nouns and terms involved in the embodiments of the present application are described, and the nouns and terms involved in the embodiments of the present application are applicable to the following explanations.

[0088] Hyperspectral Remote Sensing Images: A multidimensional information acquisition technology that combines imaging technology and spectral technology. It can not only detect the two-dimensional aggregate space information of the target, but also obtain one-dimensional spectral information, thereby generating continuous, narrow-band image data with high spectral resolution.

[0089] In the related art, hyperspectral remote sensing images record hundreds of bands for each pixel point on the electromagnetic spectrum. Different ground objects have differences in the emission, reflection and absorption of electromagnetic waves, which enables hyperspectral images to identify and detect ground objects at a fine-grained level. The rich spectral information in hyperspectral images makes them indispensable in various fields such as ecosystem measurement, mineral analysis and precision agriculture. Pixels are the basic unit of hyperspectral remote sensing image data, and the main goal of hyperspectral remote sensing image classification is to assign a unique class label to each pixel. In the prior art, various machine learning methods have been tried in the field of hyperspectral image classification, including K-nearest neighbors (K-NNs), random forests (RFs) and support vector machines (SVMs). Although these methods have achieved some success in certain cases, they generally have some limitations. First, these methods usually rely on manual feature extraction, which is not only time-consuming but also may not be accurate and efficient. Second, they fail to fully utilize the spatial and spectral characteristics of ground objects, which limits the accuracy of classification. Finally, when applied to complex datasets, the training results of these methods are often unsatisfactory. With the development of computer technology, various deep learning-based methods have been introduced in the field of hyperspectral image classification, such as stacked autoencoders (SAEs), convolutional neural networks (CNNs), transformers (Transformers) and recurrent neural networks (RNNs). Compared to traditional machine learning algorithms, these methods can automatically extract features, but they also have some limitations when dealing with hyperspectral images. Although convolutional neural networks (CNNs) perform well in image classification, they mainly focus on local feature extraction, while global information and long-range dependencies of hyperspectral images are one of the key factors for classification accuracy. CNNs have limitations in modeling these global information, which limits their performance in hyperspectral image classification. Second, although complex convolution operations can extract rich features, they also make it difficult for the model to distinguish the importance of different spectral features, which may ignore subtle features that are critical to classification. Recurrent neural networks (RNNs) are good at handling sequential data, but in the classification of hyperspectral remote sensing images, excessive reliance on the order of spectral bands may cause the gradient vanishing problem, affecting the learning effect of the model. Transformers based on multi-head self-attention mechanism can focus on the extraction of significant features when dealing with hyperspectral images, but may ignore subtle spectral differences that are crucial to improving classification performance. At the same time, hyperspectral remote sensing images contain hundreds of bands, containing a large amount of redundant information. How to extract effective features from these redundant information has a decisive influence on the performance of the model.

[0090] Therefore, the embodiment provides a hyperspectral remote sensing image classification method, which comprises the following steps: preprocessing an initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image And in the target hyperspectral remote sensing image X image The model input image is obtained by random sampling The model is fed into the image x as input. Two multi-feature fusion network modules are then used to perform refined feature extraction and fusion on the input model image x. The retention encoding network module, consisting of a simple retention layer and a feedforward network, captures the long-range dependencies between different pixels and bands in the hyperspectral remote sensing image, explores intrinsic spectral connections, and distinguishes subtle spectral differences. Finally, a global pooling layer is connected to a linear layer to obtain the target multi-feature fusion retention network model. This target multi-feature fusion retention network model is then used to predict the image to be classified, yielding the hyperspectral remote sensing image classification result.

[0091] A hyperspectral remote sensing image classification method provided in an embodiment of the present application relates to the field of image recognition technology. A hyperspectral remote sensing image classification method provided in an embodiment of the present application can be applied to a terminal, can be applied to a server, or can be software running in a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, and a car terminal, etc., but is not limited thereto; the server side can be configured as an independent physical server, or as a server cluster or distributed system composed of multiple physical servers, or as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application that implements a hyperspectral remote sensing image classification method, etc., but is not limited to the above forms.

[0092] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0093] The embodiments of the present application are specifically explained below with reference to the drawings:

[0094] Figure 1 is an optional flowchart of a hyperspectral remote sensing image classification method provided by the embodiments of the present application, Figure 1 The method in the above embodiment can include but is not limited to steps S101 to S102.

[0095] Step S101, obtaining an image to be classified;

[0096] Step S102, inputting the image to be classified into a target multi-feature fusion reservation network model to obtain a hyperspectral remote sensing image classification result.

[0097] In some embodiments, the image to be classified can be obtained first, and then input into the target multi-feature fusion reservation network model to obtain the hyperspectral remote sensing image classification result, so that the hyperspectral remote sensing image classification can be realized through the target multi-feature fusion reservation network model, and the efficiency and accuracy are improved.

[0098] In some embodiments, as shown in Figure 2 obtaining the target multi-feature fusion reservation network model can include but is not limited to steps S201 to S205.

[0099] Step S201, obtaining an initial hyperspectral remote sensing image;

[0100] Step S202, pre-processing the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image;

[0101] Step S203, randomly sampling the target hyperspectral remote sensing image according to a preset size to obtain a model input image;

[0102] Step S204, constructing an initial multi-feature fusion reservation network model;

[0103] Step S205, inputting a plurality of model input images into the initial multi-feature fusion reservation network model to train the initial multi-feature fusion reservation network model to obtain the target multi-feature fusion reservation network model.

[0104] The steps S201 to S205 shown in the embodiments of the present application obtain the target multi-feature fusion reservation network model, and improve the model precision.

[0105] In step S201 of some embodiments, the initial hyperspectral remote sensing image can be obtained through an image database. The initial hyperspectral remote sensing image can also be obtained through other ways, which are not limited. Illustratively, the obtained initial hyperspectral remote sensing image can be three-dimensional data Its bands (channels) contain useful spectral information, and the number of bands is large, where H is the height of the initial hyperspectral remote sensing image, W is the width of the initial hyperspectral remote sensing image, and C is the number of channels of the initial hyperspectral remote sensing image.

[0106] In some embodiments, in step S202, preprocessing the initial hyperspectral remote sensing image to obtain the target hyperspectral remote sensing image may include but is not limited to the following steps:

[0107] Performing principal component analysis on the initial hyperspectral remote sensing image to obtain a first image, wherein the number of channels of the first image is smaller than the number of channels of the initial hyperspectral remote sensing image;

[0108] According to the filling width, the zero-marked pixel filling method is used to fill the edge of the first image to obtain the target hyperspectral remote sensing image.

[0109] In some embodiments, preprocessing the initial hyperspectral remote sensing image may include principal component analysis and image edge filling. The initial hyperspectral remote sensing image may be subjected to principal component analysis to obtain a first image, wherein the number of channels of the first image is less than the number of channels of the initial hyperspectral remote sensing image. Exemplarily, the size of the initial hyperspectral remote sensing image may be set to 610×340×103, wherein H=610 represents height, W=340 represents width, and C=103 represents the number of channels. The principal component analysis method may be used to reduce the number of spectral bands (number of channels) in the initial hyperspectral remote sensing image X0 from C to c=30, and the first image is obtained after band reduction. Where c is the number of channels of the first image. Then, according to the filling width, the zero-marked pixel filling method is used to fill the edge of the first image to obtain the target hyperspectral remote sensing image. For example, the first image X PCA The edge pixels cannot form samples of the same size, and the spatial information (adjacent pixels in space) and spectral information (spectral information contained in the channel) of the patch sample cannot be fully utilized. You can use zero-marked pixels to fill X according to the filling width. PCA The edge of the target is obtained to obtain the final target hyperspectral remote sensing image Among them, the filling width is (P-1) / 2, h, w, c are the height, width and number of channels of the target hyperspectral remote sensing image respectively. It can be understood that the input of the model needs to be selected The size of the sample (patch), and the sample is the label of the center pixel as the input label, where p is the height and width of the model input. Further, when the required model input image is , where 9 is the height and width, and 30 is the number of channels. The center pixel is for the model input image x, and this center pixel is located at the red pixel on the 9×9 surface. The label represented by this center pixel is the model input image The label of the center pixel is shown as Figure 3 In addition, the edge pixel refers to the pixel point of the edge of the initial hyperspectral remote sensing image, and the edge pixel is filled as shown in Figure 4 The blue area is the initial hyperspectral remote sensing image, and the black area is the zero pixel point filled by the edge, and the filling width is 4. Since the edge pixel of the first image after dimension reduction cannot form a 9x9x30 patch sample, image edge filling is needed, and the height and width of the expanded image can fully utilize the spatial information and spectral information of the patch sample. More, a plurality of initial hyperspectral remote sensing images can be preprocessed to obtain corresponding target hyperspectral remote sensing images, and the plurality of target hyperspectral remote sensing images can be combined to obtain a model training data set.

[0110] In some embodiments, in step S203, the target hyperspectral remote sensing image can be randomly sampled according to a preset size to obtain a model input image. Illustratively, after the first image is edge filled to obtain the target hyperspectral remote sensing image , the preset size can be set to 9x9x30, the target hyperspectral remote sensing image can be randomly sampled to obtain a cubic model input image, and the label of the center pixel of the model input image can be used as the label of the model input image. It can be understood that using this cubic model input image as input enables the model to fully utilize the spatial information and spectral information of the hyperspectral remote sensing image.

[0111] In some embodiments, in step S204, the construction process of the initial multi-feature fusion reservation network model can include but is not limited to the following steps:

[0112] Step S301, constructing a first multi-feature fusion network, the first multi-feature fusion network being used for extracting significant features of the model input image to obtain a first multi-feature fusion network output feature map;

[0113] Step S302, constructing a second multi-feature fusion network after the first multi-feature fusion network, the second multi-feature fusion network being used for extracting subtle features of the first multi-feature fusion network output feature map to obtain a second multi-feature fusion network output feature map;

[0114] Step S303, constructing a reservation encoding network after the second multi-feature fusion network, the reservation encoding network being used for extracting spatial-spectral feature internal correlation of the second multi-feature fusion network output feature map;

[0115] Step S304, constructing a global pooling layer after the reservation encoding network;

[0116] Step S305, after the global pooling layer, a linear layer is constructed.

[0117] In some embodiments, the multi-feature fusion reservation network model adopts a structure combining a convolutional neural network (multi-feature fusion network) with a reservation encoding network. A multi-feature fusion network module with channel attention (ECA) and spatial attention (ESA) is introduced to subdivide spatial spectral features into significant spatial spectral features and subtle spatial spectral features, and perform feature fusion. A reservation encoding network module is used to explore the intrinsic correlation of spatial spectral features. The multi-feature fusion reservation network model structure is as shown in Figure 5 As shown in the figure, the first multi-feature fusion network can be constructed first, then the second multi-feature fusion network, then the reservation encoding network, and finally the global pooling layer, the linear layer and the output layer to obtain the initial multi-feature fusion reservation network model. The first multi-feature fusion network output feature map can be obtained by performing significant feature extraction on the model input image through the first multi-feature fusion network. The second multi-feature fusion network output feature map can be obtained by performing subtle feature extraction on the first multi-feature fusion network output feature map through the second multi-feature fusion network. The spatial spectral feature intrinsic correlation extraction can be performed on the second multi-feature fusion network output feature map through the reservation encoding network. After the reservation encoding network, the reservation encoding network output feature is connected with a linear layer with an input dimension of 64 and an output dimension of 9 after global pooling, to obtain the model prediction output.

[0118] In some embodiments, in step S301, the construction process of the first multi-feature fusion network can include but is not limited to the following steps:

[0119] A first convolutional layer is constructed, and the convolutional kernel size of the first convolutional layer is 3x3. The first convolutional layer is used to convolve the model input image to obtain a first feature map;

[0120] After the first convolutional layer, a first channel attention layer is constructed, which is used to reorder the first feature map according to color depth to obtain a second feature map;

[0121] After the first channel attention layer, a first partial convolution processing layer is constructed, which is used to perform partial convolution processing on the dark gray feature map in the second feature map according to a preset channel number to obtain a third feature map;

[0122] After the first partial convolution processing layer, a second convolutional layer is constructed, and the convolutional kernel size of the second convolutional layer is 1x1. The second convolutional layer is used to convolve the third feature map to obtain a fourth feature map;

[0123] After the second convolutional layer, the first batch normalization layer is constructed to perform batch normalization on the fourth feature map according to a preset activation function to obtain a first multi-feature fusion network output feature map.

[0124] In some embodiments, the first multi-feature fusion network can be constructed by first constructing the first convolutional layer, then constructing the first channel attention layer, then constructing the first partial convolution processing layer, then constructing the second convolutional layer, and finally constructing the first batch normalization layer. In the feature extraction process, the first convolutional layer can be used to convolve the input image of the model to obtain the first feature map, wherein the size of the convolution kernel of the first convolutional layer is 3x3. It can be understood that only the 3x3 convolutional layer can change the size of the feature map. In the first multi-feature fusion network, the convolutional layer with a 3x3 kernel increases the number of channels of the feature map from 30 to 40. The first convolutional layer can extract significant features. Then, according to the color depth, the first feature map is reordered by the first channel attention layer to obtain the second feature map. For example, in the first multi-feature fusion network, the channel attention ECA can reorder the first feature map according to the channel weight, so that all dark gray feature maps are placed in front of the channel dimension, and light gray feature maps are placed at the back of the channel dimension. The dark gray feature maps contain more significant information in the first and second feature maps, while the light gray feature maps are relatively subtle. Then, according to the preset number of channels, the dark gray feature maps in the second feature map are partially convolved by the first partial convolution processing layer to obtain the third feature map. It can be understood that the partial convolution processing in the spatial attention ESA first extracts significant spatial spectral feature information from the dark gray feature maps in the second feature map, and then splices the black feature map obtained by the partial convolution with the light gray part of the second feature map to generate the third feature map. Then, the third feature map is convolved by the second convolutional layer to obtain the fourth feature map, so that the features are fused, wherein the size of the convolution kernel of the second convolutional layer is 1x1. Finally, the fourth feature map is batch-normalized by the first batch normalization layer according to a preset activation function to obtain the first multi-feature fusion network output feature map.

[0125] In some embodiments, the dark gray feature maps in the second feature map are partially convolved according to the preset number of channels to obtain the third feature map, including:

[0126] According to the preset number of channels, the second feature map is divided to obtain a first divided feature map and a second divided feature map, the first divided feature map has a preset number of channels, and the number of dark gray feature maps in the first divided feature map is greater than the number of dark gray feature maps in the second divided feature map.

[0127] The first divided feature map is convolved to obtain a to-be-spliced feature map.

[0128] The to-be-spliced feature map is spliced with the second divided feature map to obtain a third feature map.

[0129] In some embodiments, the second feature map can be divided according to a preset channel number to obtain a first divided feature map and a second divided feature map. The first divided feature map has a preset number of channels, and the number of dark gray feature maps in the first divided feature map is greater than the number of dark gray feature maps in the second divided feature map. Then, the first divided feature map is convolved to obtain a to-be-spliced feature map, and finally the to-be-spliced feature map is spliced with the second divided feature map to obtain a third feature map. For example, as shown in FIG. 6, given a second feature map with a height of 9, a width of 9, and a channel number of 24, a partial convolution is performed on a selected region (i.e., a first divided feature map) of the second feature map. The selected region has a height of 9, a width of 9, and a channel number of 8, which is a dark gray feature map part sorted in the front. At the same time, in order to ensure that the spatial dimensions and the channel number of the output third feature map remain consistent with the input second feature map, the to-be-spliced feature map is spliced with the unconvoluted part (i.e., a second divided feature map with a height of 9, a width of 9, and a channel number of 16) of the second feature map in the channel dimension to obtain the third feature map, which maintains the same spatial dimensions and channel number as the second feature map. The partial convolution processing reduces the calculation redundancy and memory requirements, and is an effective feature extraction method. Figure 6

[0130] In some embodiments, in step S302, the construction process of the second multi-feature fusion network can include but is not limited to the following steps:

[0131] A third convolutional layer is constructed, the convolutional kernel size of the third convolutional layer is 3x3, and the third convolutional layer is used to convolve the first multi-feature fusion network output feature map to obtain a fifth feature map;

[0132] After the third convolutional layer, a second channel attention layer is constructed, and the second channel attention layer is used to reorder the fifth feature map according to color depth to obtain a sixth feature map;

[0133] After the second channel attention layer, a second partial convolution processing layer is constructed, and the second partial convolution processing layer is used to perform partial convolution processing on the light gray feature map in the sixth feature map according to a preset channel number to obtain a seventh feature map;

[0134] After the second partial convolution processing layer, a fourth convolutional layer is constructed, the convolutional kernel size of the fourth convolutional layer is 1x1, and the fourth convolutional layer is used to convolve the seventh feature map to obtain an eighth feature map;

[0135] ​After the fourth convolutional layer, a second batch normalization layer is constructed, which is used to perform batch normalization on the eighth feature map according to a preset activation function to obtain a second multi-feature fusion network output feature map.

[0136] In some embodiments, the third convolutional layer can be constructed first, then the second channel attention layer, the second partial convolution processing layer, the fourth convolutional layer, and finally the second batch normalization layer, to obtain the second multi-feature fusion network. In the feature extraction process, the first multi-feature fusion network output feature map can be convolved by the third convolutional layer first to obtain a fifth feature map, wherein the size of the convolution kernel of the third convolutional layer is 3x3. In the second multi-feature fusion network, the convolution kernel with a size of 3x3 increases the channel number of the feature map from 40 to 64. Then, the fifth feature map is reordered by the second channel attention layer according to the color depth to obtain a sixth feature map, and the light gray feature map in the sixth feature map is partially convolved by the second partial convolution processing layer according to the preset channel number to obtain a seventh feature map. Then, the seventh feature map is convolved by the fourth convolutional layer to obtain an eighth feature map, wherein the size of the convolution kernel of the fourth convolutional layer is 1x1. Finally, the eighth feature map is batch normalized by the second batch normalization layer according to the preset activation function to obtain the second multi-feature fusion network output feature map. It can be understood that the second multi-feature fusion network is partially similar to the first multi-feature fusion network, and the main difference lies in the partial convolution processing process in the spatial attention ESA. The partial convolution processing is performed on the second feature map after the attention ECA processing, that is, the light gray feature map with a height of 9, a width of 9, and a channel number of 8, to realize fine feature extraction and capture fine spatial spectral feature information, and the convolution kernel with a size of 1x1 is used for feature fusion in the multi-feature fusion network. The residual skip connection in the multi-feature fusion network helps to propagate feature information to deeper levels and effectively capture spatial spectral feature information.

[0137] In some embodiments, the construction process of the encoding network in step S303 can include but is not limited to the following steps:

[0138] A feature flattening layer is constructed, which is used to flatten the second multi-feature fusion network output feature map to obtain a feature sequence;

[0139] After the feature flattening layer, a simple reservation layer is constructed, which is used to perform feature extraction processing on the feature sequence to obtain a simple reservation layer output feature.

[0140] After the simple reservation layer, a feedforward network layer is constructed.

[0141] In some embodiments, the reserved encoding network (RE) consists of two key components: a simple reserved layer (SR) and a feed-forward network layer (FFN). The feature flattening layer can be built first, followed by the simple reserved layer, and then the feed-forward network layer. During feature extraction, the second multi-feature fusion network output feature map can be flattened first through the feature flattening layer to obtain a feature sequence flattening to obtain a feature sequence The feature sequence is taken as the input of the RE to learn high-level semantic features and capture the global position information of different pixels in the hyperspectral remote sensing image. Then, the feature sequence is subjected to feature extraction processing by the simple reserved layer to obtain simple reserved layer output features. Finally, the simple reserved layer output features are input into the feed-forward network layer to obtain feed-forward network layer output features. The feed-forward network layer FFN consists of a linear layer, a nonlinear activation function layer, and another linear layer in sequence. Its purpose is to learn more complex nonlinear features to improve the model fitting to the data. In addition, there is a layer normalization after the FFN to speed up the model training and reduce the risk of overfitting, thereby improving the generalization and precision of the model. More specifically, the feature extraction mathematical expression of the reserved encoding network RE can include: Y = SR(LN(X)) + X, Z = FFN(LN(Y)) + Y, RE(X) = LN(Z), where Y is the simple reserved layer output feature, SR is the simple reserved layer, LN is the layer normalization, X is the feature matrix, Z is the feed-forward network layer output feature, FFN is the feed-forward network layer, and RE is the reserved encoding network.

[0142] In some embodiments, the feature sequence is subjected to feature extraction processing to obtain simple reserved layer output features, including:

[0143] The feature sequence is subjected to matrix conversion to obtain a feature matrix;

[0144] A query matrix is calculated according to the feature matrix and a query learnable matrix;

[0145] A key matrix is calculated according to the feature matrix and a key learnable matrix;

[0146] A value matrix is calculated according to the feature matrix and a value learnable matrix;

[0147] A combined matrix of the causal mask and the exponential decay along the relative distance is calculated according to the scalar constant;

[0148] The simple reserved layer output features are calculated according to the query matrix, the key matrix, the value matrix, and the combined matrix of the causal mask and the exponential decay along the relative distance.

[0149] In some embodiments, the feature sequence can be subjected to matrix conversion first to obtain a feature matrix wherein X comprises n vectors, each vector has a dimension of d. Then a query matrix is calculated according to the feature matrix and the query learnable matrix, wherein the calculation formula of the query matrix is: Q=XW Q , wherein Q is the query matrix, X is the feature matrix, and W Q is the query learnable matrix. A key matrix is calculated according to the feature matrix and the key learnable matrix, wherein the calculation formula of the key matrix is: K=XW K , wherein K is the key matrix, and W K is the key learnable matrix. A value matrix is calculated according to the feature matrix and the value learnable matrix, wherein the calculation formula of the value matrix is: V=XW V , wherein V is the value matrix, and W V is the value learnable matrix. It can be understood that the simple reservation layer SR projects the feature matrix X into the query matrix Q, the key matrix K, and the value matrix V. The three projections can be obtained by mapping through the learnable matrix Then a combination matrix of the causal mask and the exponential decay along the relative distance is calculated according to the scalar constant, wherein the calculation formula of the combination matrix is: , wherein D nm is the combination matrix of the causal mask and the exponential decay along the relative distance, γ n-m is the scalar constant, n and m both represent integers from 0 to d-1, and d is the dimension of the vector. Finally, a simple reservation layer output feature is calculated according to the query matrix, the key matrix, the value matrix, and the combination matrix of the causal mask and the exponential decay along the relative distance, wherein the expression of the simple reservation layer output feature is: SR(X)=(QK T ⊙D)V, wherein SR(X) is the simple reservation layer output feature, K T represents the transpose matrix of K, D is the combination matrix of the causal mask and the exponential decay along the relative distance, It can be understood that the simple reservation layer SR uses D for standardization.

[0150] In some embodiments, in step S205, the plurality of model input images can be input into the initial multi-feature fusion reservation network model, so that the initial multi-feature fusion reservation network model is trained to obtain the target multi-feature fusion reservation network model. Illustratively, a certain proportion of training sample quantity can be randomly selected for training the initial multi-feature fusion reservation network model, that is, all categories are divided into training samples according to this proportion, and the rest are test samples. The training sample quantity is selected and the model is trained, and 5% of each sample in the Pavia University data set can be randomly selected as training samples for training. At the same time, the training period can be set to 100, the training batch can be set to 64, the learning rate can be set to 0.001, and the Adam optimizer can be used for optimization to obtain the target multi-feature fusion reservation network model.

[0151] In some embodiments, after obtaining the target multi-feature fusion reservation network model at the end of training, test samples can be used for testing to obtain the hyperspectral remote sensing image classification result. In the Pavia University dataset, 95% of each sample is randomly selected as a test sample for testing. The hyperspectral remote sensing image also has technical indicators for comparing and evaluating the classification result. Common evaluation indicators include Kappa coefficient (Kappa), overall classification accuracy (OA), and average classification accuracy (AA). In order to verify the method of the present embodiment, experiments are performed on the Pavia University dataset. The present embodiment is compared with the HybridSn, PiT and HiT models based on deep learning on the Pavia University dataset. The model classification accuracy comparison results are shown in Figure 7 According to the ground object number, the ground object is divided into 9 categories. In the three evaluation indicators, the present embodiment is better than the other three algorithms, and the highest accuracy is obtained in the classification and identification of multiple ground objects. In the classification result graph, the present embodiment is compared with the HybridSn, PiT and HiT models based on deep learning. The three-channel RGB image is shown in Figure 8 , the real ground object label graph is shown in Figure 9 , the HybridSn classification result graph is shown in Figure 10 , the PiT classification result graph is shown in Figure 11 , the HiT classification result graph is shown in Figure 12 , and the classification result graph of the present embodiment is shown in Figure 13 . By comparing the classification result graphs of various algorithms, the images all have different degrees of fuzzy spots. However, the classification result graph of the present embodiment is the cleanest, has the least noise, and is closest to the real ground object label graph.

[0152] In some embodiments, the overall process of training the target multi-feature fusion reservation network model is shown in Figure 14 . The hyperspectral remote sensing image dataset can be obtained first, preprocessed, and the sample size for inputting the model is selected. Then, the multi-feature fusion reservation network model is initially constructed, a certain proportion of training samples is selected for training the network model, and the target multi-feature fusion reservation network model is obtained. Further, the network performance can be tested, and the classification result is obtained.

[0153] In some embodiments, two multi-feature fusion network modules are adopted, which process a large amount of redundant information of the hyperspectral remote sensing image from the perspective of redundancy. The module combines the channel attention mechanism and the spatial attention mechanism. After 3x3 convolution, the channel attention (ECA) is used to rearrange the feature map in the channel dimension according to the weight size, and the part with large feature weight is arranged in front of the channel dimension, and the part with small weight is arranged at the back. Then, the partial convolution in the spatial attention ESA is used to perform fine feature extraction on the feature map and perform feature fusion. In the feature extraction process, only part of the feature map can be convolved, and 1x1 convolution is used for feature fusion after the feature map is spliced in the channel. The partial convolution of ESA in the first multi-feature fusion network module obtains the significant features, and then performs feature fusion. The partial convolution of ESA in the second multi-feature fusion network module obtains the subtle features, and then performs feature fusion, which effectively improves the feature extraction capability of the hyperspectral remote sensing image and the feature fusion effect. The encoding network module of the present embodiment has a simple reserved layer structure, which can effectively deal with a large amount of redundant information, avoid neglecting subtle spectral information, and effectively distinguish subtle spectral differences.

[0154] The beneficial effects of implementing the embodiments of the present application include: the present application first acquires a to-be-classified image, then inputs the to-be-classified image into a target multi-feature fusion reserved network model, and obtains a hyperspectral remote sensing image classification result, thereby realizing hyperspectral remote sensing image classification and improving efficiency and accuracy. The target multi-feature fusion reserved network model acquires an initial hyperspectral remote sensing image, pre-processes the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image, randomly samples the target hyperspectral remote sensing image according to a preset size, obtains a model input image, constructs an initial multi-feature fusion reserved network model, and finally inputs multiple model input images into the initial multi-feature fusion reserved network model to train the initial multi-feature fusion reserved network model, thereby improving model precision through the multi-feature fusion network and the reserved encoding network.

[0155] As shown in Figure 15 , the present application further provides a hyperspectral remote sensing image classification device, which comprises:

[0156] The first module 801 is used for acquiring a to-be-classified image.

[0157] The second module 802 is used for inputting the to-be-classified image into a target multi-feature fusion reserved network model to obtain a hyperspectral remote sensing image classification result.

[0158] The target multi-feature fusion reserved network model is obtained through the following steps:

[0159] An initial hyperspectral remote sensing image is acquired;

[0160] The initial hyperspectral remote sensing image is preprocessed to obtain a target hyperspectral remote sensing image;

[0161] According to a preset size, the target hyperspectral remote sensing image is randomly sampled to obtain a model input image;

[0162] An initial multi-feature fusion reservation network model is constructed;

[0163] The plurality of model input images are input into the initial multi-feature fusion reservation network model, so that the initial multi-feature fusion reservation network model is trained to obtain a target multi-feature fusion reservation network model.

[0164] The content in the above method embodiments is applicable to the device embodiments, the device embodiments specifically implement the functions same as those of the above method embodiments, and achieve the same beneficial effects as those of the above method embodiments.

[0165] As shown in Figure 16 , the present application embodiment further provides a computer device, comprising:

[0166] At least one processor 901;

[0167] At least one memory 902 for storing at least one program;

[0168] When the at least one program is executed by the at least one processor, the at least one processor implements the method shown in Figure 1 .

[0169] The content in the above method embodiments is applicable to the device embodiments, the device embodiments specifically implement the functions same as those of the above method embodiments, and achieve the same beneficial effects as those of the above method embodiments.

[0170] The preferred embodiments of the application are described above with reference to the accompanying drawings, and the scope of the right of the application embodiment is not limited by this. Any modification, equivalent replacement and improvement made by those skilled in the art without departing from the scope and essence of the application embodiment shall be within the scope of the right of the application embodiment.

Claims

1. A hyperspectral remote sensing image classification method, characterized in that: The following steps are involved: Get the image to be classified; Inputting the image to be classified into a target multi-feature fusion retention network model to obtain a hyperspectral remote sensing image classification result; The target multi-feature fusion retention network model is obtained by the following steps: Acquire initial hyperspectral remote sensing images; Preprocessing the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image; Randomly sampling the target hyperspectral remote sensing image according to a preset size to obtain a model input image; Construct an initial multi-feature fusion preservation network model; Inputting a plurality of the model input images into the initial multi-feature fusion retention network model to train the initial multi-feature fusion retention network model to obtain the target multi-feature fusion retention network model; The construction process of the initial multi-feature fusion retention network model includes: Constructing a first multi-feature fusion network, wherein the first multi-feature fusion network is used to extract significant features of the model input image to obtain a first multi-feature fusion network output feature map; After the first multi-feature fusion network, a second multi-feature fusion network is constructed, wherein the second multi-feature fusion network is used to extract subtle features from the feature map output by the first multi-feature fusion network to obtain the feature map output by the second multi-feature fusion network; After the second multi-feature fusion network, a retention coding network is constructed, wherein the retention coding network is used to extract the intrinsic correlation of spatial spectral features from the feature map output by the second multi-feature fusion network; After the retention encoding network, a global pooling layer is constructed; After the global pooling layer, a linear layer is constructed.

2. The method according to claim 1, characterized in that The preprocessing of the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image includes: Performing principal component analysis on the initial hyperspectral remote sensing image to obtain a first image, wherein the number of channels of the first image is smaller than the number of channels of the initial hyperspectral remote sensing image; According to the filling width, the first image is edge-filled using a zero-marked pixel filling method to obtain the target hyperspectral remote sensing image.

3. The method according to claim 1, characterized in that The process of constructing the first multi-feature fusion network includes: Construct the first convolutional layer, the convolution kernel size of the first convolutional layer is , the first convolutional layer is used to convolve the model input image to obtain a first feature map; After the first convolutional layer, a first channel attention layer is constructed, where the first channel attention layer is used to reorder the first feature map according to color depth to obtain a second feature map; After the first channel attention layer, construct a first partial convolution processing layer, wherein the first partial convolution processing layer is used to perform partial convolution processing on the dark gray feature map in the second feature map according to a preset number of channels to obtain a third feature map; After the first convolution processing layer, a second convolution layer is constructed. The convolution kernel size of the second convolution layer is , the second convolutional layer is used to convolve the third feature map to obtain a fourth feature map; After the second convolutional layer, a first batch of normalization layers is constructed, and the first batch of normalization layers is used to batch normalize the fourth feature map according to a preset activation function to obtain the first multi-feature fusion network output feature map.

4. The method according to claim 1, wherein The construction process of the second multi-feature fusion network includes: Construct the third convolutional layer, the convolution kernel size of the third convolutional layer is , the third convolutional layer is used to convolve the feature map output by the first multi-feature fusion network to obtain a fifth feature map; After the third convolutional layer, construct a second channel attention layer, wherein the second channel attention layer is used to reorder the fifth feature map according to color depth to obtain a sixth feature map; After the second channel attention layer, construct a second partial convolution processing layer, wherein the second partial convolution processing layer is used to perform partial convolution processing on the light gray feature map in the sixth feature map according to a preset number of channels to obtain a seventh feature map; After the second convolution processing layer, a fourth convolution layer is constructed, and the convolution kernel size of the fourth convolution layer is , the fourth convolutional layer is used to convolve the seventh feature map to obtain an eighth feature map; After the fourth convolutional layer, a second batch normalization layer is constructed, and the second batch normalization layer is used to perform batch normalization on the eighth feature map according to a preset activation function to obtain the second multi-feature fusion network output feature map.

5. The method according to claim 1, wherein The construction process of the retention coding network includes: Constructing a feature flattening layer, wherein the feature flattening layer is used to flatten the feature map output by the second multi-feature fusion network to obtain a feature sequence; After the feature flattening layer, a simple retention layer is constructed, wherein the simple retention layer is used to perform feature extraction processing on the feature sequence to obtain output features of the simple retention layer; After the simple retention layer, a feed-forward network layer is constructed.

6. The method according to claim 3, characterized in that The method of performing partial convolution processing on the dark gray feature map in the second feature map according to the preset number of channels to obtain a third feature map includes: Dividing the second feature map according to a preset number of channels to obtain a first divided feature map and a second divided feature map, wherein the first divided feature map has the preset number of channels, and the number of dark gray feature maps in the first divided feature map is greater than the number of dark gray feature maps in the second divided feature map; Convolving the first divided feature map to obtain a feature map to be spliced; The feature map to be spliced ​​is spliced ​​with the second divided feature map to obtain the third feature map.

7. The method according to claim 5, characterized in that The performing feature extraction processing on the feature sequence to obtain the output features of the simple retention layer includes: Performing matrix transformation on the feature sequence to obtain a feature matrix; Calculating a query matrix based on the feature matrix and the query learnable matrix; Calculating a bond matrix based on the feature matrix and the bond learnable matrix; Calculating a value matrix based on the feature matrix and the value learnable matrix; According to the scalar constant, the combination matrix of the causal mask and the exponential decay along the relative distance is calculated; The simple retention layer output features are calculated according to the query matrix, the key matrix, the value matrix and the combination matrix of the causal mask and the exponential decay along the relative distance.

8. A hyperspectral remote sensing image classification device, characterized in that: include: The first module is used to obtain the image to be classified; The second module is used to input the image to be classified into the target multi-feature fusion retention network model to obtain the hyperspectral remote sensing image classification result; The target multi-feature fusion retention network model is obtained by the following steps: Acquire initial hyperspectral remote sensing images; Preprocessing the initial hyperspectral remote sensing image to obtain a target hyperspectral remote sensing image; Randomly sampling the target hyperspectral remote sensing image according to a preset size to obtain a model input image; Construct an initial multi-feature fusion preservation network model; Inputting a plurality of the model input images into the initial multi-feature fusion retention network model to train the initial multi-feature fusion retention network model to obtain the target multi-feature fusion retention network model; The construction process of the initial multi-feature fusion retention network model includes: Constructing a first multi-feature fusion network, wherein the first multi-feature fusion network is used to extract significant features of the model input image to obtain a first multi-feature fusion network output feature map; After the first multi-feature fusion network, a second multi-feature fusion network is constructed, wherein the second multi-feature fusion network is used to extract subtle features from the feature map output by the first multi-feature fusion network to obtain the feature map output by the second multi-feature fusion network; After the second multi-feature fusion network, a retention coding network is constructed, wherein the retention coding network is used to extract the intrinsic correlation of spatial spectral features from the feature map output by the second multi-feature fusion network; After the retention encoding network, a global pooling layer is constructed; After the global pooling layer, a linear layer is constructed.

9. A computer device, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1 to 7.

Images