A multimodal recognition method for ship target types under local information loss conditions

Abstract

The present invention discloses a multimodal recognition method for ship target types under the condition of local information missing, and belongs to the field of data recognition technology. The method recognizes the type of ship targets through a multimodal type recognition model; the multimodal type recognition model includes a visible light image processing branch, an infrared image processing branch, a splicing module and a graph neural network; the visible light image processing branch and the infrared image processing branch respectively extract corresponding local key features from the visible light image and infrared image of the same ship target. The present invention extracts local key features from the visible light and infrared two-modal images as graph model nodes, and then uses the graph neural network to fuse the extracted information nodes, effectively overcoming the problem of missing information of a single visible light image in environments such as fog, rain and snow, and improving the accuracy of classification and recognition.

Description

Technical Field

[0001] The invention belongs to the technical field of data identification, and in particular relates to a method for identifying the type of a ship target. Background Art

[0002] Maritime ship target recognition is one of the key technologies for the development of ship intelligence and is of great significance to improving ship navigation safety. With the advancement of deep learning technology, the ship target recognition method based on neural network has achieved remarkable results. Existing methods mainly rely on visible light images. However, due to environmental factors such as clouds, rain and fog, the global characteristics of ship targets will change and local information will be missing, thus affecting the performance of the recognition algorithm.

[0003] At present, there are three main solutions for ship target type recognition under the condition of local information loss:

[0004] 1. Optimize image quality through image defogging algorithms. Related algorithms can be divided into two categories: defogging algorithms based on image enhancement (He Huan et al., Traffic image defogging method based on wavelet transform, Electronic Design Engineering, 2020) and defogging algorithms based on atmospheric scattering models (Zhang Ran, Foggy image enhancement algorithm based on fractional-order partial differential equations, Xi'an: Xi'an University of Technology, 2018). The former does not explore the causes of foggy image degradation, but makes the image more in line with the human visual system by enhancing image contrast and other methods; the latter starts from the causes of foggy image degradation and restores the image by compensating for foggy image loss.

[0005] Second, the method of using negative sample enhancement learning. For example, Yao Hongge et al. (Yao Hongge, Wang Cheng, Yu Jun, Bai Xiaojun, Li Wei, Small Target Ship Recognition in Complex Satellite Images, Journal of Remote Sensing, 2020, 24(2):116-125) added a large number of pictures containing only clouds and coastal backgrounds to the sample data set for negative sample expansion, enhancing the network model's ability to learn the features of clouds and coastal backgrounds to cope with the impact of complex sea conditions.

[0006] 3. Classification model based on local area. For example, the paper "The application of two-level attention models in deep convolutional neural network for fine-grained image classification" (Tianjun Xiao, Yichong Xu, Kuiyuan Yang, Jiaxing Zhang, YuxinPeng and Z. Zhang,, 2015, CVPR) generates candidate image blocks through a search algorithm, filters out irrelevant background, and trains object-level classification models and local-level classification models respectively. The object-level classification network DomainNet aggregates the classification results of all candidate image blocks to obtain object-level classification results; the local-level classification network clusters the convolution kernels of DomainNet into three categories based on spectral clustering, which are used as detectors for the head, torso, and object, respectively, and concatenates the filtered local area features to obtain local-level classification results through the SVM classifier; finally, the local classification results are fused with the object-level classification results to obtain the final classification results.

[0007] However, the above solutions mainly achieve classification under conditions of local information loss by enhancing global image or local image features. On the one hand, they do not consider and utilize the relationship between local features. On the other hand, they cannot solve the fundamental problem that visible light images cannot provide complete information after being disturbed by the environment, resulting in accuracy that cannot meet the requirements. Summary of the invention

[0008] The present invention proposes a multimodal recognition method for ship target types under the condition of local information loss, the purpose of which is to solve the problem of low accuracy in recognition of ship target types under the condition of local information loss.

[0009] The technical solution of the present invention is as follows:

[0010] A multi-modal recognition method for ship target types under the condition of local information loss, which recognizes the type of ship targets through a multi-modal type recognition model;

[0011] The multimodal type recognition model includes a visible light image processing branch, an infrared image processing branch, a splicing module and a graph neural network;

[0012] The visible light image processing branch and the infrared image processing branch extract corresponding local key features from the visible light image and infrared image of the same ship target respectively. The stitching module is used to stitch the two sets of local key features and obtain graph data features and adjacency matrix features based on the stitching results. The graph neural network is used to obtain recognition results based on the graph data features and the adjacency matrix features.

[0013] As a further improvement of the multimodal ship target type recognition method under the condition of local information loss: the visible light image processing branch and the infrared image processing branch have the same structure, both including a multi-scale feature learning network and an adaptive key feature learning module connected in sequence;

[0014] The multi-scale feature learning network is used to extract multi-scale features from an input image, and the adaptive key feature learning module is used to learn local key features from the multi-scale features.

[0015] As a further improvement of the multimodal recognition method for ship target types under the condition of local information loss: the multi-scale feature learning network includes a backbone network for extracting basic features and a feature aggregation network for fusing the basic features;

[0016] The backbone network extracts 5 layers of basic features , , , and ;

[0017] The feature aggregation network has 4 layers, which is similar to the basic feature , , and One-to-one correspondence, each layer of the feature aggregation network corresponds to the output , , and That is, the multi-scale features extracted by the multi-scale feature learning network.

[0018] As a further improvement of the multimodal recognition method of ship target types under the condition of local information loss, it is assumed that a certain layer in the feature aggregation network is related to the basic feature Correspondingly, this layer has nodes arranged in sequence, and the value of the first node in this layer is recorded as , the value of the last node is recorded as , the node between them is the intermediate node, and the number of convolution kernels corresponding to this layer is recorded as ;

[0019] The calculation method of each node value in the feature aggregation network is:

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] in, Indicates that the convolution kernel size is 1×1 and the number of convolution kernels is The convolution calculation, Indicates splicing, , and for The intermediate node of the corresponding layer.

[0029] As a further improvement of the multimodal ship target type recognition method under the condition of local information loss: the adaptive key feature learning module includes the output of the feature aggregation network , , and One-to-one correspondence of multi-layer polarized self-attention modules , , , , let the feature map Input to the corresponding multi-layer polarization self-attention module The local key characteristic feature map obtained is , then the local key features obtained by the adaptive key feature learning module are: ,in Represents a deconvolution operation.

[0030] As a further improvement of the multimodal recognition method for ship target types under the condition of local information loss: the multi-layer polarization self-attention module includes a channel attention branch and a spatial attention branch; the channel attention branch is used to extract the channel relationship of the input feature map, and the spatial attention branch is used to model the dependency relationship between the input feature maps; the local key characteristic feature map output by the multi-layer polarization self-attention module is the result of pixel-by-pixel addition of the output attention map of the channel attention branch and the output attention map of the spatial attention branch.

[0031] As a further improvement of the multimodal recognition method of ship target type under the condition of local information loss:

[0032] The processing process of the channel attention branch is: Assume that the feature map input by the multi-layer polarization self-attention module is , first calculate the feature map The average value of the feature map after conv(1×1) convolution on each channel, and then the average value of all channels is used to form the average feature , and then set the weights of the two fully connected layers and , weight and The initial value of is randomly generated and optimized by gradient backpropagation during training, and then the output attention map of the channel attention branch is calculated. :

[0033] ;

[0034] in, is the sigmoid function;

[0035] The processing process of the spatial attention branch is: Assume that the feature map input by the multi-layer polarization self-attention module is , first the feature map Perform several parallel convolution operations, each with a different convolution kernel size and at least one of which is a dilated convolution operation, then concatenate the results of each convolution operation, and then convolve the concatenated result and input it into the activation function Then the activation function Results and feature maps Multiply them together to get the output attention map of the spatial attention branch .

[0036] As a further improvement of the multimodal ship target type recognition method under the condition of local information loss: in the spatial attention branch, the feature map is first A convolution operation with a convolution kernel size of 3×3 and three dilated convolution operations with convolution kernel sizes of 3×3, 1×6, and 6×1 were performed; the four convolution results were concatenated and then convolved with a convolution kernel size of 3×3, and then activated by the activation function. After the feature map Multiply.

[0037] As a further improvement of the multimodal ship target type recognition method under the condition of local information loss, the activation function Defined as:

[0038] ;

[0039] in, This is a hyperparameter for adjusting the importance of key features. The larger the value, the more obvious the difference in activation effects of different key features. It is the threshold for distinguishing key features from irrelevant features. The larger the value, the fewer key features are screened.

[0040] As a further improvement of the multimodal recognition method for ship target types under the condition of local information loss: in the splicing module, the local key features of the visible light image processing branch and the local key features of the infrared image processing branch are spliced, and then expanded into a number of vectors according to channels, and the several vectors constitute the graph data feature X, and then the similarity between the vectors is calculated, and the adjacency matrix feature A is constructed according to the similarity;

[0041] Among them, Vector and Vector The similarity is:

[0042] ;

[0043] dot represents the dot product, represents the norm, is the first feature in the adjacency matrix A Row, No. Elements of a column;

[0044] The graph neural network is classified based on the graph data feature X and the adjacency matrix feature A. It includes two layers of GCNconv. After the graph data feature X and the adjacency matrix feature A pass through the two layers of GCNconv, they are activated by the softmax function to obtain the One-to-one correspondence between ship types nodes, The type corresponding to the maximum value in the nodes is the recognition result of the model.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The multi-scale feature learning network designed by the present invention extracts multi-scale features through the multi-scale feature learning network, laying the foundation for subsequent feature fusion learning, and then further realizes bottom-up and top-down bidirectional fusion between different layers through the feature aggregation network, and better fuses information through the guidance of deep semantic information on shallow detail features and the correction of shallow detail features on deep semantic information, thereby more fully extracting local information features and preventing information loss.

[0047] 2. The adaptive key feature learning module designed in the present invention includes a multi-layer polarized self-attention module (PSM). Each layer of PSM can capture long-term and large-space dependencies through channel attention and spatial attention mechanisms to better model features; among them, the introduction of dilated convolution in the spatial attention branch and the design of dilated convolution kernels of different angles in combination with the aspect ratio of the ship are conducive to the targeted extraction of features that fit the hull shape. Multi-layer PSM can perform adaptive key feature learning on features of different layers to extract local key area features of different sizes with category discrimination, providing initial relationships for the next step of advanced relationship learning.

[0048] 3. The present invention constructs a graph neural network to classify graph data constructed from images. By using local key features extracted from visible light and infrared modal images as graph model nodes, and then using the graph neural network to fuse the extracted information nodes, the high-order relationship between multi-modal key features is learned. During the training process, the graph neural network can adaptively use high-quality node information to make up for low-quality nodes, achieve information complementarity of high-order relationships, effectively overcome the problem of missing information of a single visible light image in environments such as fog, rain and snow, and improve the accuracy of classification and recognition.

[0049] 4. The multimodal ship type recognition method constructed by this method obtains local key features of different sizes through multi-scale feature learning and key feature extraction, and regards the local key features of the two modal data as graph model nodes, and then uses the graph convolutional neural network for training. This method can promote the complementarity of key information through the feature extraction of convolutional neural networks and the fusion of node information of graph neural networks, thereby reducing the impact of local information loss on recognition and improving the classification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the overall framework of the multimodal type recognition model of the present invention;

[0051] Figure 2 Schematic diagram of the structure of the multi-scale feature learning network;

[0052] Figure 3 It is a schematic diagram of the structure of the adaptive key feature learning module;

[0053] Figure 4 Schematic diagram of the structure of the graph neural network. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0055] A multimodal recognition method for ship target types under the condition of local information loss is proposed, which recognizes the type of ship targets through a multimodal type recognition model.

[0056] like Figure 1 The multimodal type recognition model includes a visible light image processing branch, an infrared image processing branch, a splicing module and a graph neural network. The visible light image processing branch and the infrared image processing branch extract corresponding local key features from the visible light image and the infrared image of the same ship target respectively. The splicing module is used to splice the two sets of local key features and obtain graph data features and adjacency matrix features based on the splicing results. The graph neural network is used to obtain recognition results based on the graph data features and the adjacency matrix features.

[0057] Specifically, the visible light image processing branch and the infrared image processing branch have the same structure, both of which include a multi-scale feature learning network and an adaptive key feature learning module connected in sequence. The multi-scale feature learning network is used to extract multi-scale features from the input image, and the adaptive key feature learning module is used to learn local key features from the multi-scale features.

[0058] like Figure 2 ,The multi-scale feature learning network includes a backbone network for extracting basic features and a feature aggregation network for fusing the basic features.

[0059] The backbone network is based on the ResNet18 network trained with the ImageNet dataset, which extracts 5 layers of basic features. , , , and . ResNet18 The features output by the residual block. In this embodiment, the basic features , , , and The feature map sizes and channel numbers are 56×56×64, 56×56×64, 28×28×128, 14×14×256, and 7×7×512, respectively.

[0060] The feature aggregation network has 4 layers, which is similar to the basic feature , , and One-to-one correspondence, let a layer in the feature aggregation network correspond to the basic feature Correspondingly, this layer has nodes arranged in sequence, and the value of the first node in this layer is recorded as , the value of the last node is recorded as , the one between them is the intermediate node (the value names are respectively , ...are arranged in alphabetical order, where There is no intermediate node in the corresponding layer), the number of convolution kernels corresponding to this layer is recorded as In this embodiment, the number of convolution kernels in the four layers of the feature aggregation network is: =64, =128, =256, =512.

[0061] The calculation method of each node value in the feature aggregation network is:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] in, Indicates that the convolution kernel size is 1×1 and the number of convolution kernels is The convolution calculation, Indicates splicing, , and for The intermediate node of the corresponding layer.

[0071] The output of the feature aggregation network , , and It is a multi-scale feature extracted by a multi-scale feature learning network. In this embodiment, , , and The scales are 56×56×64, 28×28×128, 14×14×256, and 7×7×512 respectively.

[0072] like Figure 3 The adaptive key feature learning module includes the output of the feature aggregation network , , and One-to-one multi-layer polarization self-attention module (PSM) , , , , let the feature map Input to the corresponding multi-layer polarization self-attention module The local key characteristic feature map obtained is , then the local key features obtained by the adaptive key feature learning module are: ,in , indicating that the convolution kernel size is 3, the step size is 2, and the number of convolution kernels is The number of input feature map channels The deconvolution operation is half of that.

[0073] The PSM module does not change the feature map Size, four groups The sizes are 56×56×64, 28×28×128, 14×14×256, 7×7×512, and local key features The size is 56×56×64.

[0074] Furthermore, the multi-layer polarized self-attention module includes a channel attention branch and a spatial attention branch. The channel attention branch is used to extract the channel relationship of the input feature map, and the spatial attention branch is used to model the dependency relationship between the input feature maps.

[0075] The processing process of the channel attention branch is: Assume that the feature map input by the multi-layer polarization self-attention module is , first calculate the feature map The average value of the feature map after conv(1×1) convolution on each channel, and then the average value of all channels is used to form the average feature , and then set the weights of the two fully connected layers and , weight and The initial value of is randomly generated and optimized by gradient backpropagation during training, and then the output attention map of the channel attention branch is calculated. :

[0076] ;

[0077] in, is the sigmoid function.

[0078] The processing process of the spatial attention branch is: Assume that the feature map input by the multi-layer polarization self-attention module is , first the feature map Perform several parallel convolution operations, each with a different convolution kernel size and at least one of which is a dilated convolution operation, then concatenate the results of each convolution operation, and then convolve the concatenated result and input it into the activation function Then the activation function Results and feature maps Multiply them together to get the output attention map of the spatial attention branch .

[0079] In this embodiment, considering that the length-to-width ratio of a ship is usually 6:1, the characteristic map A convolution operation with a convolution kernel size of 3×3 and three dilated convolution operations with convolution kernel sizes of 3×3, 1×6, and 6×1 were performed; the four convolution results were concatenated and then convolved with a convolution kernel size of 3×3, and then activated by the activation function. After the feature map Multiply, the processing process can be expressed as:

[0080] ;

[0081] in, represents a convolution operation with a convolution kernel size of 3×3. represents an expanded convolution operation with a convolution kernel size of 1×6, and so on.

[0082] Since the specific structure on the ship target is category-related and usually more beneficial to classification, it is expected to pay more attention to the high-response part on the feature map and ignore the less important clues that may be used for positioning or may bring interference noise. Therefore, the activation function of this method is Defined as:

[0083] ;

[0084] in, This is a hyperparameter to adjust the importance of key features. The larger the value, the more obvious the difference in activation effects of different key features. Here it is set to 1; is the threshold for distinguishing key features from irrelevant features. The larger the value, the fewer key features are screened. It is set to 0.5 here. This activation function can retain the effect of high response features, while the attention weight is less than In this way, the classifier can pay more attention to the areas with class discriminativeness and reduce the risk of overfitting and misclassification.

[0085] The local key feature map output by the multi-layer polarized self-attention module is the result of pixel-by-pixel addition of the output attention map of the channel attention branch and the output attention map of the spatial attention branch, that is, .

[0086] Furthermore, in the splicing module, the local key features of the visible light image processing branch and the local key features of the infrared image processing branch are spliced ​​and then expanded into a number of vectors by channel. The several vectors constitute the graph data feature X, and then the similarity between the vectors is calculated, and the adjacency matrix feature A is constructed based on the similarity.

[0087] Among them, Vector and Vector The similarity is:

[0088] ;

[0089] dot represents the dot product, represents the norm, is the first feature in the adjacency matrix A Row, No. Elements of a column.

[0090] In this embodiment, two local key features of size 56×56×64 are concatenated into a feature map of size 56×56×128, which is then expanded into 128 vectors of length 3136, and the resulting adjacency matrix feature size is 128×128.

[0091] Furthermore, if Figure 4 , the graph neural network is classified based on the graph data feature X and the adjacency matrix feature A. In this embodiment, the graph neural network includes two layers of GCNconv, with the number of nodes being 128 and 64 respectively. After the two layers of GCNconv, the softmax activation function is used to obtain the One-to-one correspondence between ship types nodes, The type corresponding to the maximum value in the nodes is the recognition result of the model.

[0092] Furthermore, when training the above multimodal type recognition model, the same ship target is firstly captured by infrared camera and visible light camera under conditions of rain, snow, cloud and fog. Then the ship target is cropped out of the above images and uniformly resized to 128×128×3. The cropped ship targets are then cleaned and normalized, and their categories are manually labeled. types, thereby obtaining the <visible light ship image, infrared ship image, type label> data pairs under the condition of local information loss, and then randomly dividing the training and test samples in a 7:3 ratio to construct the training set and test set.

[0093] This example uses SGD as the optimizer, the learning rate is set to 0.0002, and the batch size is set to 64. The visible light ship image and infrared ship image in the training set data pair are input into the multimodal type recognition model to obtain the recognition result , and then calculate the multi-classification cross entropy loss function:

[0094] ;

[0095] in, The first label in the data pair type A label value of 1 indicates that the group of images belongs to Type, 0 means it does not belong to Types; To identify the results Middle The probability value indicates that the model predicts that the group of images belongs to The probability of a type.

[0096] The modules of the multimodal type recognition model are optimized according to the average value of the multi-classification cross entropy loss function of this batch. The training optimization is repeated for 500 batches in total.

[0097] After the model training is completed, the visible light image and the infrared image of the same ship target captured by the visible light camera and the infrared camera are input into the multimodal type recognition model to obtain the recognition result.

[0098] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. The scope of the present invention is defined by the claims rather than the above description.

Claims

1. A multimodal recognition method for ship target types under the condition of local information loss, characterized by: The type of ship target is identified through a multi-modal type identification model; The multimodal type recognition model includes a visible light image processing branch, an infrared image processing branch, a splicing module and a graph neural network; The visible light image processing branch and the infrared image processing branch extract corresponding local key features from the visible light image and the infrared image of the same ship target respectively, the stitching module is used to stitch the two sets of local key features, and obtain graph data features and adjacency matrix features based on the stitching results, and the graph neural network is used to obtain recognition results according to the graph data features and the adjacency matrix features; The visible light image processing branch and the infrared image processing branch have the same structure, both of which include a multi-scale feature learning network and an adaptive key feature learning module connected in sequence; The multi-scale feature learning network is used to extract multi-scale features from the input image, and the adaptive key feature learning module is used to learn local key features from the multi-scale features; The multi-scale feature learning network includes a backbone network for extracting basic features and a feature aggregation network for fusing basic features; The backbone network extracts 5 layers of basic features , , , and ; The feature aggregation network has 4 layers, which is similar to the basic feature , , and One-to-one correspondence, each layer of the feature aggregation network corresponds to the output , , and That is, the multi-scale features extracted by the multi-scale feature learning network; The adaptive key feature learning module includes the output of the feature aggregation network , , and One-to-one correspondence of multi-layer polarized self-attention modules , , , , let the feature map Input to the corresponding multi-layer polarization self-attention module The local key characteristic feature map obtained is , then the local key features obtained by the adaptive key feature learning module are: ,in represents the deconvolution operation; The multi-layer polarized self-attention module includes a channel attention branch and a spatial attention branch; the channel attention branch is used to extract the channel relationship of the input feature map, and the spatial attention branch is used to model the dependency relationship between the input feature maps; the local key characteristic feature map output by the multi-layer polarized self-attention module is the result of pixel-by-pixel addition of the output attention map of the channel attention branch and the output attention map of the spatial attention branch.

2. The multi-modal identification method for ship target types under the condition of local information loss according to claim 1 is characterized in that: A layer in the feature aggregation network and the basic features Correspondingly, this layer has nodes arranged in sequence, and the value of the first node in this layer is recorded as , the value of the last node is recorded as , the node between them is the intermediate node, and the number of convolution kernels corresponding to this layer is recorded as ; The calculation method of each node value in the feature aggregation network is: ； ； ； ； ； ； ； ； in, Indicates that the convolution kernel size is 1×1 and the number of convolution kernels is The convolution calculation, Indicates splicing, , and for The intermediate node of the corresponding layer.

3. The multi-modal identification method of ship target type under the condition of local information loss according to claim 1, characterized in that: The processing process of the channel attention branch is: Assume that the feature map input by the multi-layer polarization self-attention module is , first calculate the feature map The average value of the feature map after conv(1×1) convolution on each channel, and then the average value of all channels is used to form the average feature , and then set the weights of the two fully connected layers and , weight and The initial value of is randomly generated and optimized by gradient backpropagation during training, and then the output attention map of the channel attention branch is calculated. : ； in, is the sigmoid function; The processing process of the spatial attention branch is: Assume that the feature map input by the multi-layer polarization self-attention module is , first the feature map Perform several parallel convolution operations, each with a different convolution kernel size and at least one of which is a dilated convolution operation, then concatenate the results of each convolution operation, and then convolve the concatenated result and input it into the activation function Then the activation function Results and feature maps Multiply them together to get the output attention map of the spatial attention branch .

4. The multi-modal identification method of ship target type under the condition of local information loss according to claim 3 is characterized in that: In the spatial attention branch, the feature map is first A convolution operation with a convolution kernel size of 3×3 and three dilated convolution operations with convolution kernel sizes of 3×3, 1×6, and 6×1 were performed; the four convolution results were concatenated and then convolved with a convolution kernel size of 3×3, and then activated by the activation function. After the feature map Multiply.

5. The multimodal identification method of ship target type under the condition of local information loss according to claim 4 is characterized in that: Activation Function Defined as: ； in, This is a hyperparameter for adjusting the importance of key features. The larger the value, the more obvious the difference in activation effects of different key features. It is the threshold for distinguishing key features from irrelevant features. The larger the value, the fewer key features are screened.

6. The multi-modal identification method of ship target type under the condition of local information loss according to claim 1, characterized in that: In the splicing module, the local key features of the visible light image processing branch and the local key features of the infrared image processing branch are spliced, and then expanded into a number of vectors by channel. The several vectors constitute the graph data feature X, and then the similarity between the vectors is calculated, and the adjacency matrix feature A is constructed according to the similarity; Among them, Vector and Vector The similarity is: ； dot represents the dot product, represents the norm, is the adjacency matrix feature A Row, No. Elements of a column; The graph neural network is classified based on the graph data feature X and the adjacency matrix feature A. It includes two layers of GCNconv. After the graph data feature X and the adjacency matrix feature A pass through the two layers of GCNconv, they are activated by the softmax function to obtain the One-to-one correspondence between ship types nodes, The type corresponding to the maximum value in the nodes is the recognition result of the model.

Images