Ship image classification device based on deep neural network feature coding

Abstract

The invention discloses a ship image classification device based on deep neural network feature coding. The device comprises an SAR image data acquisition module, a multi-source SAR image database module, an image encoder module, a deep residual neural network classifier module and a ship target classification display module. According to the ship image classification system, feature coding of the SAR ship image is realized by adopting the device, and the feature coding is further input into the deep residual neural network classifier module, so that the deep learning model can flexibly adapt to the diversity of the ship image, and the recognition capability of different ship types is improved. The SAR ship image recognition method based on deep learning overcomes the defects that an existing SAR ship image recognition system based on deep learning is not accurate in ship variant diversity recognition, not robust to image interference and noise and the like, and accurate classification is conducted on SAR ship images through the strong feature extraction capacity of the Transform encoder and the accurate classification capacity of the deep residual neural network.

Description

Technical Field

[0001] The present invention relates to the field of SAR ship target recognition, and particularly to a ship image classification device based on deep neural network feature encoding. Background Art

[0002] SAR radar for maritime ship target monitoring and recognition has the advantages of all-weather and full-sea area ship target detection and recognition. It is a core technology for marine safety, maritime search and rescue, illegal smuggling monitoring, situation assessment, and threat estimation, providing integrated monitoring technology support for maritime safety.

[0003] However, due to the diversity of ship variants caused by external disturbances (variants caused by changes in the ship itself, environment, and sensors, etc.) and the incompleteness of massive radar monitoring data (a large amount of data that has not been strictly manually calibrated or has errors in label calibration), it poses difficult problems for accurate ship recognition modeling. The existing technologies have unsatisfactory model recognition effects on such targets (the recognition is more difficult, but it often occurs in actual ship recognition, is more practical, and has more application scenarios).

[0004] Deep learning is a branch of machine learning. It solves complex tasks by constructing multi-layer neural networks and learning hierarchical feature representations. Deep learning models usually contain multiple hidden layers and can automatically learn features from a large amount of data, thus reducing the need for manual feature engineering and domain expertise. Ships have diverse shapes and variants in radar images. Deep learning helps capture abstract and high-level features in ship images. Deep learning models can flexibly adapt to the diversity of ship images, thereby improving the recognition ability for different ship types. SAR ship radar images may be affected by various interferences and noises. Deep learning models can learn robust representations of these interferences and improve the recognition performance under incomplete data. Summary of the Invention

[0005] The purpose of the present invention is to provide a ship image classification device based on deep neural network feature encoding to perform high-precision classification detection on SAR ship images, aiming at the deficiencies of the current SAR ship image recognition system based on deep learning, such as inaccurate recognition of ship variant diversity and non-robustness to image interference and noise.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is: A ship image classification device based on deep neural network feature encoding, comprising:

[0007] An SAR image data acquisition module for realizing active observation and data acquisition of ground ship targets;

[0008] A multi-source SAR image database module for providing diverse SAR images and ensuring the spatio-temporal consistency of the SAR images;

[0009] An image encoder module for extracting features from the input SAR image, using a feature encoder based on a deep neural network to capture and represent key features in the image, and generating a high-dimensional feature vector;

[0010] A deep residual neural network classifier module for receiving the high-dimensional feature vector output by the image encoder module and classifying ship targets using a deep residual network structure;

[0011] A ship target classification display module for visually displaying the classification results and supporting viewing the classification results through user interface interaction. The classification results include target category, confidence score, and statistical information.

[0012] Furthermore, the SAR image data acquisition module includes:

[0013] An integrated SAR sensor for acquiring SAR image data of ground ship targets;

[0014] A configuration data acquisition control unit for adjusting the working parameters of the integrated SAR sensor. The working parameters include observation angle and polarization mode;

[0015] An integrated data storage module for storing the acquired SAR image data in real time and establishing an index.

[0016] Furthermore, the multi-source SAR image database module includes:

[0017] A data acquisition and integration module, including an interface with the data source, for acquiring and integrating SAR images to form a unified multi-source SAR image database;

[0018] A data calibration and correction module for calibrating and correcting the acquired SAR images. The correction includes attitude correction and atmospheric correction;

[0019] A data storage and index management module. Among them, the data storage module includes a file system and a database; the index management module is used for quickly retrieving and querying data and supports access to the database by each module of the system.

[0020] Furthermore, the image encoder module includes the following training process:

[0021] (A1) There are m SAR ship image training data sets from different sources in the multi-source SAR image database Among them, the qth data set There are n sq Sample images and their corresponding labels Among them Denote the j-th SAR ship image in the dataset, and its corresponding label; The test target dataset contains n t labeled samples and n t << n sq , where the label y is in one-hot encoding format; Read all the training datasets in the multi-source SAR image database;

[0022] (A2) Reshape and flatten the ship image data in the training dataset to obtain where (H, W) is the resolution of the original image, C is the number of channels, (P, P) is the resolution of the segmented image, and N = HW / P 2 is the number of image segmentation blocks. The image segmentation block method is used to effectively capture the long-range dependencies in the image, enabling the deep learning model to flexibly adapt to the diversity of ship images, improving the recognition ability for different ship types, and partially overcoming the problem that traditional models are difficult to capture global image features;

[0023] (A3) Construct an image encoder model: Use the encoder of Transformer to improve the recognition performance of the model under incomplete data and partially overcome the problem that traditional models are easily affected by interference and noise in the image. A constant latent vector size D is used in all layers. The patches are flattened and mapped to the D dimension using a trainable linear projection, as shown in the following equation:

[0024]

[0025] where, z 0 represents the input of the image encoder model, x class represents the learnable special embedding, E represents the learnable linear projection mapping, and E pos represents the standard one-dimensional learnable position encoding, represents the i-th flattened image patch;

[0026] The image encoder part consists of alternating layers of multi-head self-attention MSA and MLP blocks, stacked three layers in total. The long-range dependencies in the image are further effectively captured through multi-head self-attention, enabling the deep learning model to flexibly adapt to the diversity of ship images, improving the recognition ability for different ship types, and partially overcoming the problem that traditional models are difficult to capture global image features; Layer normalization is applied before each block, and a residual connection is applied after each block. The specific operation is as follows:

[0027] z′ l = MSA(LN(z l-1 )) + zl-1 , e = 1...L (2)

[0028] z l = MLP(LN(z′ l )) + z′ l , l = 1…L (3)

[0029]

[0030] Where LN is the aforementioned LayerNorm; MLP includes two layers of linear mapping and one layer of GELU non - linear mapping; MSA represents the multi - head attention mechanism; x class The encoder output at the corresponding position As the feature encoding representation of the entire SAR ship image, for After MLP mapping and softmax normalization, the obtained y represents the class probability vector of the ship image; z l-1 Represents the image encoding output of the (l - 1)th layer, z l Represents the image encoding output of the lth layer, z′ l Represents the intermediate variable output by the lth layer;

[0031] (A4) Initialize the image encoder model network and input the training data to start training;

[0032] (A5) The training Loss uses the cross - entropy loss function, and the calculation formula is as follows:

[0033]

[0034] Where Y represents the true class label of the training data set, P represents the model output probability, N represents the total number of samples, K represents the total number of classes, y i,k Represents that the ith ship sample has the true class label k, p i,k Represents the probability that the model outputs that the ith ship sample has the class label k;

[0035] (A6) Use the Adam optimization algorithm to update the network parameters.

[0036] Furthermore, the deep residual neural network classifier module further improves the recognition performance of the model under incomplete data and partially overcomes the problem that traditional models are easily affected by interference and noise in images. It includes the following training process:

[0037] (B1) Input the training data in the multi - source SAR image database into the image encoder module to obtain the latent space encoding representation {z 1 , z 2 , …, z N}, and its corresponding class label {y 1 , y 2 ,..., y N}; N represents the total number of samples, y i represents the one-hot encoded label;

[0038] (B2) Construct a deep neural network classifier: The classifier uses a 10-layer deep residual neural network. The forward propagation formula of the l-th layer is as follows:

[0039]

[0040] Among them, represents the hidden vector representation of the i-th sample at the l-th layer. Relu represents the activation function, Relu(x) = max(x, 0), represents the MLP network, W l represents the weight vector of the MLP network at the l-th layer, b l represents the bias vector of the MLP network at the l-th layer; The final output layer uses softmax normalization to obtain the class probability vector of the SAR ship image:

[0041]

[0042] Among them, p i represents the class probability vector of the i-th SAR ship image;

[0043] (B3) The training Loss uses the cross-entropy loss function, and the calculation formula is as follows:

[0044]

[0045] Among them, Y represents the true class label of the training dataset, P represents the model output probability, N represents the total number of samples, K represents the total number of classes, y i,k represents that the i-th ship sample has the true class label k, p i,k represents the probability that the i-th ship sample output by the model has the class label k;

[0046] (B4) Use the Adam optimization algorithm to update the network parameters.

[0047] Further, the device further includes a real-time online testing module; The real-time online testing module is specifically:

[0048] (C1) Real-time collect ship images through the SAR image data collection module;

[0049] (C2) Transmit the real-time collected SAR ship images into the test dataset of the multi-source SAR image database module for storage;

[0050] (C3) Retrieve the test data set of the multi-source SAR image database module and input it into the image encoder module to obtain the latent space encoding representation of each SAR ship image in the test data set;

[0051] (C4) Input the latent space encoding representation of each SAR ship image in the test data set into the deep residual neural network classifier module to obtain the final class prediction probability vector of each SAR ship image, and then take the ship class corresponding to the maximum probability as the predicted class of the final SAR ship image;

[0052] (C5) The SAR ship image and its corresponding model prediction class information are jointly transmitted to the ship target classification display module through the bus, and their corresponding relationship is recorded and displayed;

[0053] (C6) The user obtains the class of the relevant ship through the ship target classification display module.

[0054] The beneficial effects of the present invention are as follows: 1. Aiming at the problem that traditional models are difficult to capture global image features, the present invention adopts image block division and self-attention mechanism, which can effectively capture long-range dependence relationships in images, enabling the deep learning model to flexibly adapt to the diversity of ship images and improving the recognition ability for different ship types; 2. Aiming at the problem that traditional models are easily affected by interference and noise in images, the present invention adopts a Transformer encoder and a deep residual neural network to improve the recognition performance of the model under incomplete data. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a schematic structural diagram of the system of the present invention;

[0056] Figure 2 is a schematic structural diagram of the Transformer encoder in the image encoder module of the present invention;

[0057] Figure 3 is a schematic structural diagram of the residual connection in the deep residual neural network classifier module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The present invention will be specifically described below with reference to the drawings.

[0059] As Figure 1 shown, the embodiment of the present invention provides a ship image classification device based on deep neural network feature encoding, which consists of a SAR image data acquisition module, a multi-source SAR image database module, an image encoder module, a deep residual neural network classifier module, and a ship target classification display module. The ship image classification system of the present invention can perform high-precision classification on ship targets in SAR images.

[0060] An embodiment of the present invention also provides a ship image classification method based on deep neural network feature encoding based on the above device, and the specific steps are as follows:

[0061] Step 1: Real-time collect ship images through the SAR image data acquisition module.

[0062] Step 2: Input the real-time collected SAR ship images into the test data set of the multi-source SAR image database module for storage.

[0063] Step 3: Take out the test data set of the multi-source SAR image database module and input it into the trained image encoder to obtain the latent space encoding representation of each SAR ship image in the test data set.

[0064] Step 4: Input the latent space encoding representation of each SAR ship image in the test data set into the trained deep residual neural network classifier to obtain the final class prediction probability vector of each SAR ship image, and then take the ship class corresponding to the maximum probability as the predicted class of the final SAR ship image.

[0065] Step 5: The SAR ship image and its corresponding model prediction class information are jointly transmitted to the ship target classification display module through the bus, and their corresponding relationship is recorded and displayed.

[0066] Step 6: The user obtains the class of the relevant ship through the ship target classification display module and uses this as a basis for the control station to take further actions.

[0067] In summary, after steps such as ship image acquisition, database storage, latent space representation of the image encoder, and deep residual neural network classification, the ship image classification system can achieve a high-precision classification task for ship targets in SAR images.

[0068] Next, the implementation manner and function of each module in a ship image classification device based on deep neural network feature encoding of the present invention will be described in more specific detail.

[0069] The SAR image data acquisition module includes the following components:

[0070] Integrate an advanced SAR sensor for actively acquiring SAR image data of ground ship targets. This sensor has the capabilities of high resolution, wide bandwidth, and all-weather observation to ensure the acquisition of SAR images with high-quality information;

[0071] Configure a data acquisition control unit for adjusting the working parameters of the SAR sensor, including observation angle, polarization mode, etc., to meet different environmental and mission requirements;

[0072] The integrated data storage module is responsible for storing the collected SAR image data in real time and establishing an index for convenient query of the subsequent multi-source SAR image database.

[0073] The role of the SAR image data acquisition module is to achieve active observation and data acquisition of ground ship targets. By using a high-performance SAR sensor, the system can obtain SAR images with detailed information under various environmental conditions, providing a high-quality data basis for subsequent image classification and target recognition.

[0074] The multi-source SAR image database module includes the following components:

[0075] Data acquisition and integration module: Through interfaces with multiple SAR sensors, satellites, and other data sources, it realizes the acquisition and integration of SAR image data from different sources. This module is responsible for obtaining and integrating SAR images from different platforms and times to form a unified multi-source SAR image database.

[0076] Data calibration and correction module: Introduce a data calibration and correction module to accurately calibrate and correct the collected SAR images to ensure the consistency of the data in space and time. This includes processing such as attitude correction and atmospheric correction to improve image quality and reduce errors.

[0077] Data storage and index management: Establish an efficient data storage structure, including a file system and a database, to accommodate large-scale SAR image data. At the same time, through index management, it realizes fast retrieval and query of data, supporting efficient access to the database by other modules of the system.

[0078] The role of the described multi-source SAR image database module is as follows: Diverse data support: Collect SAR images from different sensors and platforms to provide diverse data for the system and enhance the classification ability of ship targets; Spatiotemporal consistency: Through calibration and correction, ensure that the SAR images in the database are consistent in space and time, improving the comparability and accuracy of the data; Efficient retrieval: Establish an effective data storage and index management mechanism, enabling the system to quickly retrieve and obtain the required SAR images, enhancing the real-time performance and response speed of the system; Expandability: Support dynamic update and expansion to accommodate new SAR data sources, ensuring that the system remains stable and efficient in the face of growing data volumes. The multi-source SAR image database module provides rich and spatiotemporally consistent SAR image data for the entire system, laying a solid foundation for subsequent data preprocessing, feature extraction, and target classification steps.

[0079] An image encoder module based on Vision Transformer is used to encode the features of SAR ship images, capture the abstract and high-level features in the ship images, and improve the recognition ability of different ship types. Refer to Figure 2 , the construction and training process of the image encoder module includes the following steps:

[0080] Step A1: There are m SAR ship image training datasets from different sources in the multi-source SAR image database Among them, the qth dataset has n sq sample images and their corresponding labels Among them represents the jth SAR ship image in the dataset, represents its corresponding label; in the test target dataset there are n t labeled samples and n t <<n sq , and the label y is in one-hot encoding format. Read all the training datasets in the multi-source SAR image database.

[0081] Step A2: Reshape and flatten the ship image data in the training dataset to obtain where (H,W) is the resolution of the original image, C is the number of channels, (P,P) is the resolution of the segmented image, and N = HW / P 2 is the number of image segmentation blocks. The long-range dependencies in the image are effectively captured through the image segmentation blocks, enabling the deep learning model to flexibly adapt to the diversity of ship images, improving the recognition ability of different ship types, and partially overcoming the problem that traditional models are difficult to capture global image features.

[0082] Step A3: Build an image encoder model. The image encoder uses the encoder part of Transformer to improve the recognition performance of the model under incomplete data and partially overcome the problem that traditional models are easily affected by interference and noise in the image. A constant latent vector size D is used in all layers. Therefore, the patches are flattened and projected onto the D dimension using a trainable linear projection, as shown in the following equation:

[0083]

[0084] where, z 0 is the input of the image encoder model, x class is a learnable special embedding, E is a learnable linear projection mapping, and E pos is a standard one-dimensional learnable position encoding, Represents the flattened i-th image patch.

[0085] The image encoder part consists of alternating layers of multi-head self-attention (MSA) and MLP blocks, stacked three layers in total. Through multi-head self-attention, it can further effectively capture long-range dependencies in the image, enabling the deep learning model to flexibly adapt to the diversity of ship images, improving the recognition ability for different ship types, and partially overcoming the problem that traditional models are difficult to capture global image features. Layer normalization (LN) is applied before each block, and a residual connection is applied after each block. The specific operation is as follows:

[0086] z′ l = MSA(LN(z l-1 )) + z l-1 , l = 1...L (2)

[0087] z l = MLP(LN(z′ l )) + z′ l , l = 1...L (3)

[0088]

[0089] Among them, MLP contains two layers of linear mapping and one layer of GELU non-linear mapping. MSA is the multi-head attention mechanism. x class The encoder output at the corresponding position As the feature encoding representation of the entire SAR ship image, after performing MLP mapping and softmax normalization, the obtained y is the class probability vector of the ship image; z l-1 represents the image encoding output of the (l - 1)-th layer, z l represents the image encoding output of the l-th layer, and z′ l is the intermediate variable output of the l-th layer.

[0090] Step A4: Initialize the image encoder model network and start training by inputting training data.

[0091] Step A5: The training Loss uses the cross-entropy loss function, and the calculation formula is as follows:

[0092]

[0093] Among them, Y is the true class label of the training dataset, P is the model output probability, N is the total number of samples, K is the total number of classes, y i,k means that the i-th ship sample has the true class label k, and p i,k means the probability that the i-th ship sample output by the model has the class label k.

[0094] Step A6: Use the Adam optimization algorithm to update network parameters.

[0095] The construction and training process of the deep residual neural network classifier module has the following steps:

[0096] Step B1: Take out the training data from the multi-source SAR image database and input it into the trained image encoder to obtain the latent space encoding representation {z 1 , z 2 , …, z N} of each SAR ship image and its corresponding class label {y 1 , y 2 , …, y N}, where N represents the total number of samples, y i is a one-hot encoded label;

[0097] Step B2: Construct a deep neural network classifier. Refer to Figure 3 , the classifier uses a 10-layer deep residual neural network, and the forward propagation formula of the l-th layer is as follows:

[0098]

[0099] Among them, represents the latent vector representation of the i-th sample at the l-th layer, Relu is the activation function, Relu(x) = max(x, 0), F represents the MLP network, W l is the weight vector of the l-th layer MLP network, b l is the bias vector of the l-th layer MLP network; the final output layer uses softmax normalization to obtain the class probability vector of the SAR ship image:

[0100]

[0101] Among them, p i represents the class probability vector of the i-th SAR ship image;

[0102] Step B3: The training Loss uses the cross-entropy loss function, and the calculation formula is as follows:

[0103]

[0104] Among them, Y is the true class label of the training dataset, P is the model output probability, N is the total number of samples, K is the total number of classes, y i,k means that the i-th ship sample has the true class label k, and p i,k means the probability that the i-th ship sample output by the model has the class label k.

[0105] Step B4: Use the Adam optimization algorithm to update the network parameters.

[0106] During the online testing process, the test dataset of the multi-source SAR image database module is taken out and input into the trained image encoder to obtain the latent space encoding representation of each SAR ship image in the test dataset. Then, the latent space encoding representation of each SAR ship image in the test dataset is input into the trained deep residual neural network classifier to obtain the final class prediction probability vector of each SAR ship image. The ship class corresponding to the maximum probability is taken as the predicted class of the final SAR ship image.

[0107] The above embodiments are used to explain the present invention rather than limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A ship image classification device based on deep neural network feature coding, characterized in that: include: SAR image data acquisition module, used to realize active observation and data acquisition of ground ship targets; A multi-source SAR image database module, used to provide diversified SAR images and ensure the temporal and spatial consistency of the SAR images; The image encoder module is used to extract features from the input SAR image, and uses a feature encoder based on a deep neural network to capture and represent the key features in the image and generate a high-dimensional feature vector; A deep residual neural network classifier module is used to receive the high-dimensional feature vector output by the image encoder module and classify the ship target using a deep residual network structure; The ship target classification display module is used to visualize the classification results and support interactive viewing of the classification results through the user interface. The classification results include target category, confidence score and statistical information.

2. A ship image classification device based on deep neural network feature coding according to claim 1, characterized in that: The SAR image data acquisition module comprises: Integrated SAR sensor to obtain SAR image data of ground ship targets; A data acquisition control unit is configured to adjust the operating parameters of the integrated SAR sensor, wherein the operating parameters include an observation angle and a polarization mode; The integrated data storage module is used to store the collected SAR image data in real time and create indexes.

3. The ship image classification device based on deep neural network feature coding according to claim 1 is characterized in that: The multi-source SAR image database module includes: Data acquisition and integration module, including interfaces with data sources, used to acquire and integrate SAR images to form a unified multi-source SAR image database; A data calibration and correction module is used to calibrate and correct the collected SAR images, wherein the correction includes attitude correction and atmospheric correction; Data storage and index management module, where the data storage module includes a file system and a database; the index management module is used to quickly retrieve and query data and supports the access of various modules of the system to the database.

4. The ship image classification device based on deep neural network feature coding according to claim 1 is characterized in that: The image encoder module includes the following training process: (A1) In the multi-source SAR image database, there are m SAR ship image training datasets from different sources. The qth data set There are n sq Sample images and their corresponding labels in, represents the jth SAR ship image in the dataset, Indicates its corresponding label; test target dataset There is n t Labeled samples And n t <<n sq , the label y is in one-hot encoding format; read all training data sets in the multi-source SAR image database; (A2) Ship image data in the training dataset Reshape and flatten to get Where (H, W) represents the resolution of the original image, C represents the number of channels, (P, P) represents the resolution of the segmented image, and N = HW / P 2 Indicates the number of image segments; (A3) Construct an image encoder model: Use a Transformer encoder with a constant latent vector size D in all layers, flatten the patch and map it to D dimensions using a trainable linear projection, as shown in the following equation: Among them, z0 represents the input of the image encoder model, x class represents a learnable special embedding, E represents a learnable linear projection mapping, and E pos represents a standard one-dimensional learnable positional encoding, represents the i-th image patch after flattening; The image encoder part consists of alternating layers of multi-head self-attention MSA and MLP blocks, stacked in three layers; layer normalization is applied before each block, and residual connection is applied after each block. The specific operation is as follows: z′ l =MSA(LN(z l-1 ))+z l-1 ,l=1...L (2) With l =MLP(LN(z′ l ))+z′ l ,l=1...L (3) Among them, LN is the layer normalization; MLP contains two layers of linear mapping and one layer of GELU nonlinear mapping; MSA represents the multi-head attention mechanism; x class Encoder output corresponding to position As the feature coding representation of the entire SAR ship image, The y obtained after MLP mapping and softmax normalization represents the category probability vector of the ship image; z l-1 represents the image encoding output of the l-1th layer, z l represents the output of the image encoding at layer l, z′ l Represents the intermediate variable of the output of the lth layer; (A4) Initialize the image encoder model network and input training data to start training; (A5) The training loss uses the cross entropy loss function, and the calculation formula is as follows: Among them, Y represents the true category label of the training data set, P represents the model output probability, N represents the total number of samples, K represents the total number of categories, and y i,k Indicates that the i-th ship sample has the true category label k, p i,k Represents the probability that the i-th ship sample output by the model has the category label k; (A6) Adam optimization algorithm is used to update network parameters.

5. The ship image classification device based on deep neural network feature coding according to claim 1 is characterized in that: The deep residual neural network classifier module includes the following training process: (B1) Input the training data in the multi-source SAR image database into the image encoder module to obtain the latent space encoding representation {z1,z2,...,z N }, and its corresponding category labels {y1,y2,...,y N }, N represents the total number of samples, y i Represents a one-hot encoded label; (B2) Constructing a deep neural network classifier: The classifier uses a 10-layer deep residual neural network, where the forward propagation formula of the lth layer is as follows: in, represents the hidden vector representation of the i-th sample in the l-th layer, represents the hidden vector representation of the i-th sample in the l+1 layer, Relu represents the activation function, Relu(x)=max(x,0), represents the MLP network, W l represents the weight vector of the l-th layer MLP network, b l Represents the bias vector of the l-th layer MLP network; the final output layer uses softmax normalization to obtain the category probability vector of the SAR ship image: Among them, p i represents the class probability vector of the i-th SAR ship image, Represents the hidden vector representation of the i-th sample at the 10th layer; (B3) The training loss uses the cross entropy loss function, and the calculation formula is as follows: Among them, Y represents the true category label of the training data set, P represents the model output probability, N represents the total number of samples, K represents the total number of categories, and y i,k Indicates that the i-th ship sample has the true category label k, p i,k Represents the probability that the i-th ship sample output by the model has the category label k; (B4) Adam optimization algorithm is used to update network parameters.

6. The ship image classification device based on deep neural network feature coding according to claim 1 is characterized in that: The device also includes a real-time online testing module; The real-time online test module is specifically: (C1) Real-time acquisition of ship images through the SAR image data acquisition module; (C2) transferring the real-time collected SAR ship images to the test data set of the multi-source SAR image database module for storage; (C3) taking out the test data set of the multi-source SAR image database module, inputting it into the image encoder module, and obtaining the latent space encoding representation of each SAR ship image in the test data set; (C4) Inputting the latent space encoding representation of each SAR ship image in the test data set into the deep residual neural network classifier module to obtain the final category prediction probability vector of each SAR ship image, and then taking the ship category corresponding to the maximum probability as the final prediction category of the SAR ship image; (C5) The SAR ship image and its corresponding model prediction category information are transmitted to the ship target classification display module through the bus, and their corresponding relationship is recorded and displayed; (C6) The user obtains the category of the relevant ship through the ship target classification display module.

Images