Hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method

Abstract

The invention relates to a hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method. The method comprises the following steps: acquiring a data set of remote sensing semantic segmentation; designing a hybrid architecture expert network as a feature encoder based on the prior parameters; constructing a heterogeneous coding dynamic search space, and learning and fusing multi-granularity visual features of hybrid architecture experts through adaptive search; constructing a coarse-fine granularity representation search space, and fusing multi-granularity visual features through adaptive search learning; a bottom-to-top step-by-step up-sampling strategy is adopted, and the multi-granularity visual features are restored to the original resolution of the image from coarse to fine through adaptive search learning; and outputting a semantic segmentation drawing result, and evaluating the performance of the network by adopting precision evaluation. According to the remote sensing image semantic segmentation method, the remote sensing image semantic segmentation network with richer expression of various features is constructed by fusing multiple architectures, mixing expert visual features and fusing multi-granularity visual features, and the adaptability to the multi-scale features of remote sensing ground features is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image semantic segmentation network design, and in particular to a hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method. Background Art

[0002] Land use and land cover mapping is a crucial task in geographic research and has been widely used in a variety of fields, including monitoring global environmental change, ecological and climate change, urban planning, and resource surveys. With the advancement of remote sensing technology, the resolution of remote sensing images has become increasingly higher. In particular, with their rich spatial details and latent semantic content, remote sensing images have become the primary means of monitoring large-scale land use and land cover.

[0003] Semantic segmentation networks based on deep learning mainly rely on the following two design methods:

[0004] (1) Artificial network design based on prior knowledge

[0005] Artificial network design based on prior knowledge relies on researchers’ prior knowledge and experience, and its mainstream design methods include: single architecture design and multi-architecture hybrid design.

[0006] 1) Single-architecture design methods typically rely on a single network to design the corresponding network. Mainstream networks include: Convolutional Neural Network (CNN), Transformer, Mamba, etc. These network frameworks each have their own advantages and disadvantages in visual feature expression. For example, Convolutional Neural Network has a strong ability to learn local spatial features; the Transformer architecture enhances its ability to capture global information in the image through its self-attention mechanism; and the Mamba architecture has a more balanced visual expression.

[0007] 2) Multi-architecture hybrid design approaches enhance visual representation capabilities by integrating the aforementioned single networks. For example, a convolutional neural network and a Transformer network can be combined. Typical network architectures include FT-UnetFormer. By integrating different architectures, the advantages of each can be further combined while addressing their own shortcomings.

[0008] (2) Automated network design based on network architecture search

[0009] Automated network design based on network architecture search can effectively address the waste of manpower and computing resources caused by manual network design by automatically searching for the optimal neural network architecture. Gradient-based algorithms, currently the mainstream, can quickly and effectively adjust network architecture parameters. Furthermore, they are less dependent on computing resources, helping to find a more precise optimal architecture.

[0010] In summary, both of the above methods have certain deficiencies and defects:

[0011] (1) For artificial network design based on prior knowledge, the single-architecture design method is limited by its own design characteristics, and often has both advantages and disadvantages, which further limits the further improvement of its generalization ability. Compared with the multi-architecture hybrid design method, it is limited by the experience of artificial design, which often manifests itself in insufficient feature richness of multi-architecture network fusion and insufficient exploration of the potential of its feature expression.

[0012] (2) For automated network design based on network architecture search, since multi-architecture collaboration requires an increase in the search space, it often consumes a lot of computing resources and prolongs the training time. However, current research lacks a reasonable search solution for multi-architecture network fusion. Summary of the Invention

[0013] In order to solve the above technical problems, this application provides a hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method, which enhances the generalization ability of remote sensing ground object semantic segmentation by fusing hybrid architecture expert features and multi-granularity visual features.

[0014] This application provides a hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method, which adopts the following technical solutions:

[0015] A hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method includes the following steps:

[0016] S1, obtain a remote sensing semantic segmentation dataset and divide the dataset into a training set for network training, a validation set for network parameter adjustment, and a test set for network evaluation;

[0017] S2, a hybrid architecture expert network is designed as a feature encoder based on prior parameters;

[0018] S3, constructs a dynamic search space for heterogeneous encoding and learns to fuse multi-granular visual features of hybrid architecture experts through adaptive search;

[0019] S4, constructing a coarse-fine representation search space and fusing coarse-fine visual features through adaptive search learning;

[0020] S5 adopts a bottom-up step-by-step upsampling strategy to restore multi-granularity visual features from coarse to fine to the original image resolution through adaptive search learning;

[0021] S6 outputs the semantic segmentation mapping results and uses accuracy evaluation to evaluate the performance of the network.

[0022] Preferably, the S1 includes the following:

[0023] Obtain an open source or self-made remote sensing semantic segmentation dataset. The dataset is divided into a training set for network training, a validation set for network hyperparameter adjustment, and a test set for network accuracy evaluation according to a preset ratio. The dataset includes remote sensing images and labeled images.

[0024] Preferably, said S2 comprises the following steps:

[0025] S21. Select a hybrid architecture expert network, where the network based on three semantic segmentation network architectures is selected as the hybrid architecture expert, namely: 1) Convolutional neural network expert: Select DeepLapV3+ or PSPNet or ConvNeXt, denoted as C Expert1 , C Expert2 ,…,C ExpertN ; 2) Transformer network expert: select Segformer or SwinTransformer or PoolFormer, denoted as T Expert1 , T Expert2 ,…,T ExpertN ; 3) Mamba network expert, denoted as M Expert1 , M Expert2 ,…,M ExpertN ;

[0026] S22, based on prior evaluation, selects the hybrid architecture expert representative. For the hybrid architecture expert network, based on prior training knowledge, the optimal network is selected in each architecture as the hybrid architecture expert representative:

[0027] C Expert =Ψ(C Expert1 ,C Expert2 ,…,C ExpertN );

[0028] T Expert =Ψ(T Expert1 ,T Expert2 ,…,T ExpertN );

[0029] M Expert =Ψ(M Expert1 ,M Expert2 ,…,M ExpertN );

[0030] Among them, C Expert 、T Expert 、M Expert The best convolutional neural network, Transformer network, and Mamba network expert representatives are selected, respectively. Ψ(·) represents the network optimization strategy based on prior pre-training knowledge, that is, the model parameters with excellent training on open source or self-made datasets are used as the representative through network evaluation to obtain the highest accuracy network.

[0031] S23, hybrid architecture expert visual feature encoding, uses the optimal hybrid architecture expert representative selected in S22 as the feature encoder of the architecture to extract the visual features of the input image and generate multi-granularity visual features one by one;

[0032] C1, C2, C3, C4 = C Expert (X);

[0033] T1, T2, T3, T4 = T Expert (X);

[0034] M1,M2,M3,M4=M Exper t(X);

[0035] Where X is the input image, C Expert (·), T Expert (·), M Expert (·) are the convolutional neural network representatives, Transformer network representatives and Mamba network representatives selected from S22, respectively. i , T i and M i are the multi-granularity features obtained, i∈{1,2,3,4}.

[0036] Preferably, said S3 comprises the following steps:

[0037] S31, constructs a dynamic search space for heterogeneous encoding. For the same-scale visual features of hybrid architecture experts, a directed acyclic graph structure is constructed as the basic cell unit, and a two-layer node layer is constructed as the search space to adaptively fuse the features of different expert networks;

[0038] S32, build a hybrid architecture expert interaction module, the formula is:

[0039]

[0040] Among them, i, k∈{1,2,3,4}, and i, k are used to represent the index values of features of different scales; It is a dynamic convolution layer, which includes 1 dynamic convolution layer and 1 normalization layer. Used to enhance feature C i , T i , M i Adaptability to the mid-ground scale; is the learnable contribution weight, and the formula is:

[0041]

[0042] Among them, A is the input feature, softmax(·) is the softmax function used to calculate the contribution probability, and the TopK algorithm is used to select only the top k networks and give them contribution weights.

[0043] Preferably, said S4 comprises the following steps:

[0044] S41, builds a coarse-to-fine-grained representation search space. For the multi-granularity features fused by hybrid architecture experts, a directed acyclic graph structure is constructed as the basic cell unit, and a multi-layer node is constructed as the search space for fusing multi-granularity features:

[0045] S42, build a coarse-fine feature dynamic enhancement module, and transform the current input feature S i As the main feature, and the remaining granularity features as auxiliary features, that is:

[0046] D i =Fusion((S a ,S b ,S c ),S i );

[0047] Among them, Fusion(·) is a coarse-fine feature dynamic enhancement module, where S i is the selected main feature, S i ∈{S1,S2,S3,S4},S a ,S b ,S c For other remaining features, it includes the following steps:

[0048] S421, firstly, the auxiliary features are sampled to the same scale as the main features through a dynamic learnable parameter, the formula is:

[0049] Among them, S i ∈{S1,S2,S3,S4}; DConv(·) represents a dynamic convolution layer, BNonv(·) represents a normalization layer, DownConv(·) is a downsampling convolution layer, Conv(·) is a 1×1 convolution layer, and UpConv(·) represents an upsampling convolution layer;

[0050] S422, through self-learning search for coarse-fine feature fusion, enhance the multi-granularity representation capability of auxiliary features. The process is as follows:

[0051] Among them, θ(·) is the learnable weight contribution rate, and the formula is:

[0052]

[0053] Among them, A is the input feature, softmax(·) is the softmax function used to calculate the contribution probability, and the TopK algorithm is used to select only the top k networks to give contribution weights;

[0054] S423, based on the auxiliary features, the self-attention mechanism is used to focus on and fuse the region of interest of the main features. The process is as follows:

[0055]

[0056] Among them, Linear(·) is a linear transformation layer, c is the number of feature channels, softmax(·) is the softmax function, and Linear(·) is a linear layer.

[0057] Preferably, the S5 includes:

[0058] D4=Fusion((S1,S2,S3),S4);

[0059] D3=Fusion((S1,S2,D4),S3);

[0060] D2=Fusion((S1,D3,D4),S2);

[0061] D1=Fusion((S2,D3,D4),S1);

[0062] Wherein, Fusion(·) is executed according to S42.

[0063] Preferably, the S6 includes:

[0064] D = convup(D1);

[0065] Among them, convup(·) includes an upsampling convolution layer and a softmax layer to output the probability of each category.

[0066] In summary, the present invention has the following beneficial technical effects:

[0067] 1. By fusing multi-architecture and hybrid expert visual features as well as multi-granularity visual features, this paper constructs a remote sensing image semantic segmentation network with richer feature expressions and enhances its adaptability to the multi-scale characteristics of remote sensing objects.

[0068] 2. The present invention integrates the advantages of multiple architecture networks through adaptive search learning, further enhancing the richness of feature expression, thereby effectively suppressing the inherent defects of visual feature expression of each architecture network expert, further focusing on the network optimization of their respective advantages, and at the same time calculating the effective gain of each architecture through the hybrid expert network, suppressing noise, improving the accuracy of semantic segmentation, and enhancing the adaptability to the multi-scale characteristics of remote sensing objects.

[0069] 3. The present invention improves the scientific nature of network design through adaptive search learning, avoids the limitations of manual design, further reduces invalid parameters, thereby reducing the consumption of computing resources and improving the generalization and reasoning capabilities of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flow chart in an embodiment of the present invention.

[0071] Figure 2 Schematic diagram of the remote sensing image semantic segmentation result given in an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further described in detail below with reference to the accompanying drawings.

[0073] The embodiment of the present invention discloses a hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method.

[0074] Reference Figure 1 ,The hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method includes the following steps:

[0075] S1, obtain a remote sensing semantic segmentation dataset and divide the dataset into a training set for network training, a validation set for network parameter adjustment, and a test set for network evaluation;

[0076] Specifically include:

[0077] Obtain an open-source or self-made remote sensing semantic segmentation dataset, and divide the dataset into training set, validation set, and test set according to the preset ratio;

[0078] The dataset includes remote sensing images and labeled images, and the images are cut into 512×512 sizes. The training set is used for network training, the validation set is used for hyperparameter adjustment, and the test set is used for network accuracy evaluation. This embodiment uses the LoveDA remote sensing semantic segmentation dataset, which comes from three different cities and covers urban and rural scenes. The dataset contains 5,987 images with a high resolution of 0.3 meters, each of which has a size of 1024×1024 pixels. The images are divided into training, validation and test sets, which include 2,522, 1,669 and 1,796 images respectively. These images are labeled into seven land feature types: buildings, roads, water bodies, wastelands, forests, agricultural land and background.

[0079] S2, a hybrid architecture expert network is designed as a feature encoder based on prior parameters;

[0080] The specific steps include:

[0081] S21. Select a hybrid architecture expert network, where the network based on three semantic segmentation network architectures is selected as the hybrid architecture expert, namely: 1) Convolutional neural network expert: Select DeepLapV3+ or PSPNet or ConvNeXt, denoted as C Expert1 ,C Expert2 ,…,C ExpertN ; 2) Transformer network expert: select Segformer or SwinTransformer or PoolFormer, denoted as T Expert1 ,T Expert2 ,…,T ExpertN ; 3) Mamba network expert, denoted as M Expert1 ,M Expert2 ,…,M ExpertN ;

[0082] S22, based on prior evaluation, select the hybrid architecture expert representative. For the hybrid architecture expert network, based on prior training knowledge, select the best network in each architecture as the hybrid architecture expert representative: C Expert =Ψ(C Expert1 ,C Expert2 ,…,C ExpertN );T Expert =Ψ(T Expert1 ,T Expert2 ,…,T ExpertN );M Expert =Ψ(M Expert1 ,M Expert2 ,…,M ExpertN );

[0083] Among them, C Expert 、T Expert 、MExpert The best convolutional neural network, Transformer network, and Mamba network expert representatives are selected, respectively. Ψ(·) represents the network optimization strategy based on prior pre-training knowledge, that is, the model parameters with excellent training on open source or self-made datasets are used as the representative through network evaluation to obtain the highest accuracy network.

[0084] Multiple networks of the same architecture are selected as backbones, and a unified decoder is selected. Pre-training is performed on a semantic segmentation dataset. Using prior knowledge from accuracy evaluation, the generalization and semantic segmentation capabilities of these networks for the corresponding dataset are determined. The optimal network for a given architecture is selected using multiple metrics. Based on this strategy, the optimal network among networks of different architectures is selected and used as a feature extractor by freezing the parameters in the backbone.

[0085] The above strategy has the following advantages:

[0086] 1) Through network optimization, the network with the best performance is selected, which can better utilize the architecture's ability to express features;

[0087] 2) Through prior pre-training, the prior network parameters are obtained, which reduces the process of network parameter adjustment. By freezing the network, the memory consumption of network training is reduced;

[0088] 3) By freezing the effective parameters in Backbone, the network is more inclined to adjust the gain of architecture fusion and reduce the process of feature parameter adjustment.

[0089] S23, hybrid architecture expert visual feature encoding, uses the optimal hybrid architecture expert representative selected in S22 as the feature encoder of the architecture, extracts the visual features of the input image, and generates multi-granularity visual features one by one; C1, C2, C3, C4 = C Expert (X); T1, T2, T3, T4 = T Expert (X); M1, M2, M3, M4 = M Exper t(X);

[0090] Where X is the input image, C Expert (·), T Expert (·), M Expert (·) are the convolutional neural network representatives, Transformer network representatives and Mamba network representatives selected from S22, respectively. i , T i and M i are the multi-granularity features obtained, i∈{1,2,3,4}.

[0091] S3, constructs a dynamic search space for heterogeneous encoding and learns to fuse multi-granular visual features of hybrid architecture experts through adaptive search;

[0092] The specific steps include:

[0093] S31, constructs a dynamic search space for heterogeneous encoding. For the same-scale visual features of hybrid architecture experts, a directed acyclic graph structure is constructed as the basic cell unit, and a two-layer node layer is constructed as the search space to adaptively fuse the features of different expert networks;

[0094] S32, builds a hybrid architecture expert interaction module, which focuses on and integrates the advantages of visual representation of different hybrid architecture expert architectures by building a robust and effective network structure through self-learning search in the above heterogeneous encoding dynamic search space. Its formula is:

[0095]

[0096] Among them, i, k∈{1,2,3,4}, and i, k are used to represent the index values of features of different scales; It is a dynamic convolution layer, which includes 1 dynamic convolution layer and 1 normalization layer. Used to enhance feature C i , T i , M i ω(·) is the learnable contribution weight. By quantifying the importance of the expert features of different hybrid architectures, the contribution weight parameter is adaptively learned and sorted based on this value. The visual features of different experts are further weighted and fused. The formula is:

[0097]

[0098] Among them, A is the input feature, softmax(·) is the softmax function used to calculate the contribution probability, and the TopK algorithm is used to select only the top k networks and give them contribution weights.

[0099] S4, constructs a coarse-fine grain representation search space, learns to fuse coarse-fine grain visual features through adaptive search, and sets the main feature S i By constructing a coarse-to-fine granularity representation search space, adaptively learning the effective gain of other granularity auxiliary features for the main feature, specifically including the following steps:

[0100] S41, builds a coarse-to-fine-grained representation search space. For the multi-granularity features fused by hybrid architecture experts, a directed acyclic graph structure is constructed as the basic cell unit, and a multi-layer node is constructed as the search space for fusing multi-granularity features:

[0101] S42, build a coarse-fine feature dynamic enhancement module, and transform the current input feature S i As the main feature, and the remaining granularity features as auxiliary features, that is:

[0102] D i =Fusion((S a ,S b ,S c ),S i );

[0103] Among them, Fusion(·) is a coarse-fine feature dynamic enhancement module, where S i is the selected main feature, S i ∈{S1,S2,S3,S4},S a ,S b ,S c For other remaining features, it includes the following steps:

[0104] The features of the hybrid architecture experts of the same scale that are interactively fused by the hybrid architecture experts are further enhanced by searching for multi-granularity features from coarse to fine through self-learning;

[0105] S421, firstly, the auxiliary features are sampled to the same scale as the main features through a dynamic learnable parameter, the formula is:

[0106] Among them, S i ∈{S1,S2,S3,S4}; DConv(·) represents a dynamic convolution layer, BNonv(·) represents a normalization layer, DownConv(·) is a downsampling convolution layer, Conv(·) is a 1×1 convolution layer, and UpConv(·) represents an upsampling convolution layer;

[0107] S422, through self-learning search for coarse-fine feature fusion, enhance the multi-granularity representation capability of auxiliary features. The process is as follows:

[0108] Where θ(·) is the learnable weight contribution rate. By quantifying the importance of features at different granularities, we adaptively learn the contribution weight parameter and sort based on this value to further weightedly fuse visual features at different granularities. The formula is:

[0109]

[0110] Among them, A is the input feature, softmax(·) is the softmax function used to calculate the contribution probability, and the TopK algorithm is used to select only the top k networks and give them contribution weights.

[0111] S423, based on the auxiliary features, the self-attention mechanism is used to focus on and fuse the region of interest of the main features. The process is as follows:

[0112] Among them, Linear(·) is a linear transformation layer, c is the number of feature channels, softmax(·) is the softmax function, and Linear(·) is a linear layer.

[0113] S5 adopts a bottom-up step-by-step upsampling strategy to restore multi-granularity visual features from coarse to fine to the original image resolution through adaptive search learning;

[0114] The specific steps include:

[0115] A bottom-up progressive upsampling strategy is adopted, and the expressive power of features is further enhanced by progressive upsampling of the fused multi-granularity features. By multi-granular fusion of semantic information and detail information from coarse to fine granularity, the complementarity of features is continuously enhanced, and the multi-granularity visual features from coarse to fine granularity are restored to the original resolution of the image through adaptive search learning. The process is as follows:

[0116] D4=Fusion((S1,S2,S3),S4);

[0117] D3=Fusion((S1,S2,D4),S3);

[0118] D2=Fusion((S1,D3,D4),S2);

[0119] D1=Fusion((S2,D3,D4),S1);

[0120] Wherein, Fusion(·) is executed according to S42.

[0121] S6 outputs the semantic segmentation mapping results and uses accuracy evaluation to evaluate the performance of the network.

[0122] The output formula is:

[0123] D = convup(D1);

[0124] Among them, convup(·) includes an upsampling convolution layer and a softmax layer to output the probability of each category.

[0125] S7, compare the performance of the output semantic segmentation mapping results with the current deep learning network.

[0126] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method, characterized by: The following steps are involved: S1, obtain a remote sensing semantic segmentation dataset and divide the dataset into a training set for network training, a validation set for network parameter adjustment, and a test set for network evaluation; S2, a hybrid architecture expert network is designed as a feature encoder based on prior parameters; S3, constructs a dynamic search space for heterogeneous encoding and learns to fuse multi-granular visual features of hybrid architecture experts through adaptive search; S4, constructing a coarse-fine representation search space and fusing coarse-fine visual features through adaptive search learning; S5 adopts a bottom-up step-by-step upsampling strategy to restore multi-granularity visual features from coarse to fine to the original image resolution through adaptive search learning; S6 outputs the semantic segmentation mapping results and uses accuracy evaluation to evaluate the performance of the network.

2. A hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method, characterized by: The S1 includes the following: Obtain an open source or self-made remote sensing semantic segmentation dataset. The dataset is divided into a training set for network training, a validation set for network hyperparameter adjustment, and a test set for network accuracy evaluation according to a preset ratio. The dataset includes remote sensing images and labeled images.

3. A hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method, characterized by: The S2 comprises the following steps: S21. Select a hybrid architecture expert network, where the network based on three semantic segmentation network architectures is selected as the hybrid architecture expert, namely: 1) Convolutional neural network expert: Select DeepLapV3+ or PSPNet or ConvNeXt, denoted as C Expert1 , C Expert2 ,…,C ExpertN ; 2) Transformer network expert: select Segformer or SwinTransformer or PoolFormer, denoted as T Expert1 , T Expert2 ,…,T ExpertN ; 3) Mamba network expert, denoted as M Expert1 , M Expert2 ,…,M ExpertN ; S22, based on prior evaluation, selects the hybrid architecture expert representative. For the hybrid architecture expert network, based on prior training knowledge, the optimal network is selected in each architecture as the hybrid architecture expert representative: C Expert =Ψ(C Expert1 ,C Expert2 ,…,C ExpertN ); T Expert =Ψ(T Expert1 ,T Expert2 ,…,T ExpertN ); M Expert =Ψ(M Expert1 ,M Expert2 ,…,M ExpertN ); Among them, C Expert 、T Expert 、M Expert The best convolutional neural network, Transformer network, and Mamba network expert representatives are selected, respectively. Ψ(·) represents the network optimization strategy based on prior pre-training knowledge, that is, the model parameters with excellent training on open source or self-made datasets are used as the representative through network evaluation to obtain the highest accuracy network. S23, hybrid architecture expert visual feature encoding, uses the optimal hybrid architecture expert representative selected in S22 as the feature encoder of the architecture, extracts the visual features of the input image, and generates multi-granularity visual features one by one, namely: C1,C2,C3,C4=C Expert (X); T1,T2,T3,T4=T Expert (X); M1,M2,M3,M4=M Exper t(X); Where X is the input image, C Expert (·), T Expert (·), M Expert (·) are the convolutional neural network representatives, Transformer network representatives and Mamba network representatives selected from S22, respectively. i , T i and M i are the multi-granularity features obtained, i∈{1,2,3,4}.

4. The hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method according to claim 1 is characterized in that: The S3 includes the following steps: S31, constructs a dynamic search space for heterogeneous encoding. For the same-scale visual features of hybrid architecture experts, a directed acyclic graph structure is constructed as the basic cell unit, and a two-layer node layer is constructed as the search space to adaptively fuse the features of different expert networks; S32, build a hybrid architecture expert interaction module, the formula is: Among them, i, k∈{1,2,3,4}, and i, k are used to represent the index values of features of different scales; It is a dynamic convolution layer, which includes 1 dynamic convolution layer and 1 normalization layer. Used to enhance feature C i , T i , M i Adaptability to the mid-ground scale; is the learnable contribution weight, and the formula is: Among them, A is the input feature, softmax(·) is the softmax function used to calculate the contribution probability, and the TopK algorithm is used to select only the top k networks and give them contribution weights.

5. The hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method according to claim 1, characterized in that: The S4 comprises the following steps: S41, builds a coarse-to-fine-grained representation search space. For the multi-granularity features fused by hybrid architecture experts, a directed acyclic graph structure is constructed as the basic cell unit, and a multi-layer node is constructed as the search space for fusing multi-granularity features: S42, build a coarse-fine feature dynamic enhancement module, and transform the current input feature S i As the main feature, and the remaining granularity features as auxiliary features, that is: D i =Fusion((S a ,S b ,S c ),S i ); Among them, Fusion(·) is a coarse-fine feature dynamic enhancement module, where S i is the selected main feature, S i ∈{S1,S2,S3,S4},S a ,S b ,S c For the remaining features, it includes the following steps: S421, firstly, the auxiliary features are sampled to the same scale as the main features through a dynamic learnable parameter, the formula is: Among them, S i ∈{S1,S2,S3,S4}; DConv(·) represents a dynamic convolution layer, BNonv(·) represents a normalization layer, DownConv(·) is a downsampling convolution layer, Conv(·) is a 1×1 convolution layer, and UpConv(·) represents an upsampling convolution layer; S422, through self-learning search for coarse-fine feature fusion, enhance the multi-granularity representation capability of auxiliary features. The process is as follows: Among them, θ(·) is the learnable weight contribution rate, and the formula is: Among them, A is the input feature, softmax(·) is the softmax function used to calculate the contribution probability, and the TopK algorithm is used to select only the top k networks to give contribution weights; S423, based on the auxiliary features, the self-attention mechanism is used to focus on and fuse the region of interest of the main features. The process is as follows: Among them, Linear(·) is a linear transformation layer, c is the number of feature channels, softmax(·) is the softmax function, and Linear(·) is a linear layer.

6. The hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method according to claim 5, characterized in that: The S5 includes: D4=Fusion((S1,S2,S3),S4); D3=Fusion((S1,S2,D4),S3); D2=Fusion((S1,D3,D4),S2); D1=Fusion((S2,D3,D4),S1); Wherein, Fusion(·) is executed according to S42.

7. The hybrid architecture expert-driven remote sensing semantic segmentation self-learning network search method according to claim 1, characterized in that: The S6 includes: D = convup(D1); Among them, convup(·) includes an upsampling convolution layer and a softmax layer to output the probability of each category.

Images