Ship full type recognition method and system in a disturbed environment based on a progressive course

Abstract

This invention relates to a method for identifying all types of ships in disturbed environments based on a progressive curriculum, comprising: acquiring images of ships to be identified; and identifying the images of ships to be identified using a pre-trained ship identification model based on a progressive curriculum, thereby obtaining identification results. When training the ship identification model based on a progressive curriculum, a difficulty generator and scheduler are used to facilitate providing training data to the model with gradually increasing difficulty gradients, making it more suitable for real-world marine scenarios and improving its robustness and generalization ability, effectively handling noise and variations in the data. Simultaneously, the use of the concept of few-shot learning allows this method to improve generalization ability with a small number of training categories, greatly enhancing the model's generalization ability and practicality.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for identifying all types of ships in disturbed environments based on a progressive curriculum. Background Technology

[0002] Currently, deep learning-based ship identification models are primarily based on three main strategies: target recognition, target detection, and semantic segmentation. On one hand, these approaches present all monitoring instances randomly to the model. Furthermore, the marine environment is dynamic, high-dimensional, and complex in combination. The presence of various factors such as fog, clouds, rain, currents, and clutter further complicates the detection process, neglecting the complexity of the data samples and the current situation. This inevitably slows down the model's learning progress, leading to low robustness and low generalization performance. Secondly, based on traditional training methods, improving generalization relies on a large number of samples for new categories. However, in reality, labeling data is costly, and there may be situations where data for a certain category is scarce, which greatly reduces the model's scalability and applicability. Summary of the Invention

[0003] The purpose of this invention is to at least address one of the shortcomings of the prior art by providing a method and system for identifying all types of ships in disturbed environments based on a progressive curriculum.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] Specifically, a method for identifying all types of ships under disturbed environments based on a progressive curriculum is proposed, including the following:

[0006] Acquire images of the vessel to be identified;

[0007] The image of the ship to be identified is identified by a pre-trained, progressively-learned, all-type ship recognition model, and the recognition result is obtained.

[0008] Specifically, the ship type identification model based on a progressive curriculum is trained in the following way:

[0009] Construct a training set T, a support set S, and a query set Q for ship images;

[0010] The training set T is divided into n independent subsets, and the subsets are preprocessed to generate n course subsets with progressively increasing difficulty gradients.

[0011] Difficulty scheduling is performed on the n course subsets. When the training loss value of the previous stage tends to converge, the training course subset of the next stage is used.

[0012] A ternary feature multidimensional loss module is constructed, which includes a feature extraction unit and a multidimensional loss function unit. The data of the i-th course subset is input into the feature extraction unit, and the ternary branch obtains the feature vectors of anchor samples, positive samples and negative samples. The vector difference between each group of samples is obtained through the multidimensional loss function unit and used as the loss value for this iteration for training.

[0013] Backpropagation updates the parameters, and the feature extraction unit in the ternary feature multidimensional loss module is trained to obtain the trained feature extraction unit.

[0014] In the prediction phase, the support set S and the query set Q are input to the trained feature extraction unit to obtain feature vectors of a specific dimension. Based on the feature vectors of the specific dimension, image classification is performed on each query and support sample.

[0015] Furthermore, specifically, the subset is preprocessed by random cropping, noise amplification, and brightness adjustment to generate n course subsets T with progressively increasing difficulty gradients. n .

[0016] Furthermore, specifically, the number of course subsets is 6, i.e., n=6. The preprocessing process is as follows:

[0017] The initial training curriculum is generated by randomly cropping the data as follows.

[0018] x crop =f crop (x)

[0019] f crop =ramdom_crop(x, size)

[0020] Where x represents the original image, and x crop Indicates that by applying f crop The image is obtained by processing the random cropping function, and size represents the size of the image after cropping.

[0021] By adding noise interference to the initial course as follows, a mid-term training course is generated.

[0022] x noisy (i, j, k) = (0, with probability of p)

[0023] Where, x noisy (i, j, k) represents the pixel value of the j-th row, k-th column, and i-th channel after adding noise, and p represents the probability value of adding noise, which is used to control the degree of noise interference added to the image;

[0024] By adjusting the brightness of the mid-term courses as follows, the final training course can be generated.

[0025] x bright (i, j, k) = {x i，j，k +Δ, 0 < x i，j，k +Δ≤255}

[0026] Where, x bright (i, j, k) represents the value of the i-th channel in the j-th row and k-th column after brightness adjustment, Δ represents the brightness adjustment value, and x j，j，k This represents the pixel value of the i-th channel in the j-th row and k-th column of the image.

[0027] Furthermore, specifically, difficulty scheduling is performed on the n subsets of courses, including the following:

[0028] The training course subset used is adjusted in real time based on the training progress. When the training loss value of the previous stage tends to converge, the training course subset for the next stage begins to be used, as shown in the following formula.

[0029]

[0030] Where t represents the t-th iteration, ε represents the convergence threshold of accuracy; A(t) function represents the accuracy after the t-th iteration, T0 represents the convergence of accuracy calculation from this iteration onwards, when the number of iterations exceeds T0, the accuracy change A(t)-A(t-1) after each iteration is less than the pre-set threshold ε, and this is considered the end of this course segment.

[0031] Furthermore, the data from the i-th course subset is input into the feature extraction unit, and the ternary branch obtains the feature vectors of the anchor sample, positive sample, and negative sample, including:

[0032]

[0033] Where n is the index of the course subset, F(·θ) r ) is a convolutional neural network model, where F(·) and θ r These are the network mapping function and parameters of the convolutional neural network model, respectively; X set Let seti be the vector of features extracted, where i takes the value T. n ,S,Q.

[0034] Furthermore, specifically, the multi-dimensional loss function unit is constructed based on the multi-dimensional triplet fusion loss function, which is expressed as follows:

[0035] Loss total =[αLoss Cosine +βLoss Euclidean+γLoss Minkowski +δLoss Mahalanobis +···] + +Margin

[0036]

[0037]

[0038]

[0039]

[0040] Where the sample is x, f() is the mapping function, and f(δ) ij =∑ m ∑ n (δ) i+m ×w m,n δ can be If the size of the entire training set is N, then the input (a, p, n) for each triplet is... The Euclidean distance metric between Anchor and Positive samples. This is the Euclidean distance metric between Anchor and Negative samples. The cosine distance metric between Anchor and Positive samples. The cosine distance metric between Anchor and Negative samples;

[0041] The Minkowski distance is a metric for the distance between Anchor and Positive samples. The Minkowski distance metric between Anchor and Negative samples; The Manhattan distance metric between Anchor and Positive samples. The Manhattan distance metric is used between Anchor and Negative samples; α, β, γ, δ... are different weight parameters used to balance the influence between different distance metrics; Margin is... and The minimum interval between the distances, [ ] + This indicates that when the value within [] is greater than zero, that value is taken as the loss.

[0042] Furthermore, specifically, the backpropagation updates parameters, including:

[0043] By summing and subtracting the feature vectors of each branch, and using the sum as the network loss value for backpropagation, we can obtain the gradient information of the input image based on the loss value. Using the loss function Loss total After inverting the gradient of the values, the obtained data are for three ship image samples. gradient information, This indicates the differentiation operation.

[0044] Furthermore, specifically, the process of completing image classification includes the following:

[0045] The support set S and query set Q are input into the trained feature extraction unit. The feature vectors of the support set S and query set Q are matched, and the pair with the highest similarity is selected as the predicted category label to complete image classification. The matching is performed using the following formula:

[0046]

[0047]

[0048] y i =argmax(p i )

[0049] Where λ, μ, ν, etc. are different weighting parameters used to balance the influence between different distance metrics; p i For S i and Q i The similarity value between samples is used as the confidence score for each class, y i The category of the group with the highest cosine similarity among all groups is taken as the Q. i Predicted class label for the sample.

[0050] This invention also proposes a vessel identification device for all types under disturbed environments based on a progressive curriculum, comprising the following:

[0051] The data acquisition module is capable of acquiring images of the vessel to be identified;

[0052] The data recognition module is used to recognize the image of the ship to be identified by a pre-trained, progressive course-based ship all-type recognition model, and obtain the recognition result.

[0053] The pre-trained, progressively structured, ship-wide identification model is obtained through the following modules.

[0054] The dataset construction module is used to construct the training set T, support set S, and query set Q of ship images.

[0055] The difficulty generation module includes a random cropping unit, a noise amplification unit, and a brightness adjustment unit, which are used to divide the training set T into n independent subsets and preprocess the subsets to generate n course subsets with progressively increasing difficulty gradients.

[0056] The difficulty scheduling module is used to adjust the training course subset in real time according to the training situation. When the training loss value of the previous stage tends to converge, the training course subset of the next stage will be used.

[0057] The ternary feature multi-dimensional loss module includes a feature extraction unit and a multi-dimensional loss function unit. It is used to input the data of the i-th course subset into the feature extraction unit, the ternary branch to obtain the feature vectors of anchor samples, positive samples and negative samples, and the multi-dimensional loss function unit to obtain the vector difference between each group of samples, which is used as the loss value for this iteration for training.

[0058] The feature extraction unit training module is used for backpropagation to update parameters and train the feature extraction unit in the ternary feature multidimensional loss module to obtain the trained feature extraction unit.

[0059] The multi-dimensional measurement module is used in the prediction stage to input the support set S and the query set Q into the trained feature extraction unit to obtain feature vectors of specific dimensions, and to complete image classification for each query and support sample based on the feature vectors of specific dimensions.

[0060] A computer-readable storage medium storing a computer program that, when executed by a processor, implements steps such as those of a method for identifying all types of ships in a disturbed environment based on a progressive curriculum.

[0061] The beneficial effects of this invention are as follows:

[0062] The present invention provides a method for identifying all types of ships in disturbed environments based on a progressive curriculum. Utilizing a difficulty generator and scheduler, it facilitates the provision of training data to the model with gradually increasing difficulty gradients, making it more suitable for real-world marine scenarios and improving its robustness and generalization ability, effectively handling noise and variations in the data. Simultaneously, by employing the concept of few-shot learning, this method can enhance generalization ability with a small number of training categories, greatly improving the model's generalization capability and practicality. Attached Figure Description

[0063] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments shown in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar output voltages. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings:

[0064] Figure 1 The flowchart shown is a process for identifying all types of ships under disturbed environments based on a progressive curriculum, according to the present invention.

[0065] Figure 2 The diagram shows the training flowchart of the ship identification model based on the progressive course in the perturbed environment identification method of the present invention.

[0066] Figure 3 The diagram shown is a schematic representation of the core structure of the ship identification system under disturbed conditions based on a progressive curriculum, according to the present invention.

[0067] Figure 4 The diagram shown is a schematic representation of the matching results of a trial method for the ship identification method and system based on progressive courses under disturbed environments according to the present invention. Detailed Implementation

[0068] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.

[0069] This invention utilizes the PyTorch framework and employs a large amount of ship data across seven categories on a self-constructed ship dataset for sample metric differential learning training. The network learning and training framework uses PyTorch, and the feature extraction structure of the neural network model adopts the ResNet50 structure. An ASGD optimizer with an initial learning rate of 0.001 is used, the ship category sample image size is 224*224, and training is conducted for 30 epochs. During training, the support set consists of 5 (10-way) samples per class, and the query set consists of 200 (10-way) samples per class.

[0070] In all embodiments of this invention, training begins from scratch. First, the ship sample image features of each network layer are obtained through forward propagation. The sum of these features is then used as the loss function value. This loss function value is then used for backpropagation to obtain the gradient information of the input ship sample image based on the loss function values ​​of each layer's image features. This improves the recognition type coverage of the comparison network.

[0071] Specifically, refer to Figure 1 as well as Figure 2 Example 1: This invention proposes a method for identifying all types of ships under disturbed environments based on a progressive curriculum, including the following:

[0072] Step 110: Obtain an image of the vessel to be identified;

[0073] Step 120: Recognize the image of the ship to be identified using a pre-trained ship all-type recognition model based on progressive courses, and obtain the recognition result;

[0074] Specifically, the ship type identification model based on a progressive curriculum is trained in the following way:

[0075] Step 210: Construct the training set T, support set S, and query set Q for ship images;

[0076] Step 220: Divide the training set T into n independent subsets, and preprocess the subsets to generate n course subsets with progressively increasing difficulty gradients;

[0077] Step 230: Perform difficulty scheduling on the n course subsets. When the training loss value of the previous stage tends to converge, the training course subset of the next stage is used.

[0078] Step 240: Construct a ternary feature multidimensional loss module. The ternary feature multidimensional loss module includes a feature extraction unit and a multidimensional loss function unit. Input the data of the i-th course subset into the feature extraction unit. The ternary branch obtains the feature vectors of the anchor sample, positive sample and negative sample. The multidimensional loss function unit obtains the vector difference between each group of samples and uses it as the loss value for this iteration for training.

[0079] Step 250: Backpropagate to update parameters and train the feature extraction unit in the ternary feature multidimensional loss module to obtain the trained feature extraction unit;

[0080] Step 260: In the prediction phase, the support set S and the query set Q are input into the trained feature extraction unit to obtain feature vectors of a specific dimension. Based on the feature vectors of the specific dimension, image classification is performed on each query and support sample.

[0081] In this embodiment, a difficulty generator and scheduler are used to provide training data to the model with gradually increasing difficulty gradients, making it more suitable for real-world ocean scenarios and improving its robustness and generalization ability, effectively handling noise and variations in the data. Simultaneously, the use of the concept of few-shot learning allows this method to improve generalization ability with training data from a small number of classes, greatly enhancing the model's generalization ability and practicality.

[0082] In one implementation, a ship image training set T (7524 samples in 7 categories), a support set S (50 samples in 10 categories), and a query set Q (2000 samples in 10 categories) are constructed by the user. The categories of the support set and the query set have not appeared in the training set.

[0083] For a dataset of ship images, x i For the i-th ship image, y i ∈{1,2,…,C t} represents the category label, N=7524 represents the number of ship images, and C t =10 represents the number of ship categories; simultaneously, identification is performed through the combined use of the following preferred implementation methods to obtain the following... Figure 4 The recognition results shown demonstrate that the perturbed environment-based all-category recognition method of this invention, using a progressive curriculum strategy, leverages a difficulty generator and scheduler to provide training data to the model with gradually increasing difficulty gradients. This approach is more suitable for real-world marine scenarios, improving robustness and generalization ability, and effectively handling noise and variations in the data. Furthermore, the use of few-shot learning allows this method to enhance generalization ability with a limited number of training categories, significantly improving the model's generalization capability and practicality.

[0084] In a preferred embodiment of the present invention, the subset is preprocessed by random cropping, noise amplification, and brightness adjustment to generate n course subsets T with progressively increasing difficulty levels. n .

[0085] In a preferred embodiment of the present invention, the number of course subsets is 6, i.e., n=6, and the preprocessing process is as follows.

[0086] The initial training curriculum is generated by randomly cropping the data as follows.

[0087] x crop =f crop (x)

[0088] f crop =ramdom_crop(x, size)

[0089] Where x represents the original image, and xcrop Indicates that by applying f crop The image is obtained by processing the random cropping function, and size represents the size of the image after cropping.

[0090] By adding noise interference to the initial course as follows, a mid-term training course is generated.

[0091] x noisy (i, j, k) = (0, with probability of p)

[0092] Where, x noisy (i, j, k) represents the pixel value of the j-th row, k-th column, and i-th channel after adding noise, and p represents the probability value of adding noise, which is used to control the degree of noise interference added to the image;

[0093] By adjusting the brightness of the mid-term courses as follows, the final training course can be generated.

[0094] x bright (i, j, k) = {x i，j，k +Δ, 0 < x i，j，k +Δ≤255}

[0095] Where, x bright (i, j, k) represents the value of the i-th channel in the j-th row and k-th column after brightness adjustment, Δ represents the brightness adjustment value, and x j，j，k This represents the pixel value of the i-th channel in the j-th row and k-th column of the image.

[0096] In a preferred embodiment of the present invention, specifically, difficulty scheduling is performed on the n course subsets, including the following:

[0097] The training course subset used is adjusted in real time based on the training progress. When the training loss value of the previous stage tends to converge, the training course subset for the next stage begins to be used, as shown in the following formula.

[0098]

[0099] Where t represents the t-th iteration, ε represents the convergence threshold of accuracy; A(t) function represents the accuracy after the t-th iteration, T0 represents the convergence of accuracy calculation from this iteration onwards, when the number of iterations exceeds T0, the accuracy change A(t)-A(t-1) after each iteration is less than the pre-set threshold ε, and this is considered the end of this course segment.

[0100] In a preferred embodiment of the present invention, the data of the i-th course subset is input into the feature extraction unit, and the ternary branch obtains the feature vectors of anchor point samples, positive sample samples, and negative sample samples, including...

[0101]

[0102] Where n is the index of the course subset, F(·θ) r ) is a convolutional neural network model, where F(·) and θ r These are the network mapping function and parameters of the convolutional neural network model, respectively; X set Let seti be the vector of features extracted, where i takes the value T. n ,S,Q.

[0103] In a preferred embodiment of the present invention, the multi-dimensional loss function unit is specifically constructed based on the multi-dimensional triplet fusion loss function, which is expressed as follows:

[0104] Loss total =[αLoss Cosine +βLoss Euclidean +γLoss Minkowski +δLoss Mahalanobis +···] + +Margin

[0105]

[0106]

[0107]

[0108]

[0109] The omitted content refers to: the product of different weight parameters (factors) and different distance metrics, where the sample is x, w is the weight of the convolution kernel, i, j are the dimensions of the extracted information, m, n are the dimensions of the convolution kernel, f(x) is the mapping function, and f(δ) is the distance between the samples. ij =∑ m ∑ n (δ) i+m ×w m,n δ can be If the size of the entire training set is N, then the input (a, p, n) for each triplet is... The Euclidean distance metric between Anchor and Positive samples. This is the Euclidean distance metric between Anchor and Negative samples. The cosine distance metric between Anchor and Positive samples. The cosine distance metric between Anchor and Negative samples; The Minkowski distance is a metric for the distance between Anchor and Positive samples. The Minkowski distance metric between Anchor and Negative samples; The Manhattan distance metric between Anchor and Positive samples. The Manhattan distance metric is used between Anchor and Negative samples; α, β, γ, δ... are different weight parameters used to balance the influence between different distance metrics; Margin is... and The minimum interval between the distances, [ ] + This indicates that when the value within [] is greater than zero, that value is taken as the loss.

[0110] In a preferred embodiment of the present invention, specifically, the backpropagation update parameters include:

[0111] By summing and subtracting the feature vectors of each branch, and using the sum as the network loss value for backpropagation, we can obtain the gradient information of the input image based on the loss value. Using the loss function Loss total After inverting the gradient of the values, the obtained data are for three ship image samples. gradient information, This indicates the differentiation operation.

[0112] In a preferred embodiment of the present invention, the process of completing image classification specifically includes the following:

[0113] The support set S and query set Q are input into the trained feature extraction unit. The feature vectors of the support set S and query set Q are matched, and the pair with the highest similarity is selected as the predicted category label to complete image classification. The matching is performed using the following formula:

[0114]

[0115]

[0116] y i =argmax(p i )

[0117] Where λ, μ, ν, etc. are different weighting parameters used to balance the influence between different distance metrics; p i For S i and Q i The similarity value between samples is used as the confidence score for each class, y i The category of the group with the highest cosine similarity among all groups is taken as the Q. i Predicted class label for the sample.

[0118] Combination Figure 3 The present invention also proposes a ship identification device for all types under disturbed environments based on a progressive curriculum, comprising the following:

[0119] The data acquisition module is capable of acquiring images of the vessel to be identified;

[0120] The data recognition module is used to recognize the image of the ship to be identified by a pre-trained, progressive course-based ship all-type recognition model, and obtain the recognition result.

[0121] The pre-trained, progressively structured, ship-wide identification model is obtained through the following modules.

[0122] The dataset construction module is used to construct the training set T, support set S, and query set Q of ship images.

[0123] The difficulty generation module includes a random cropping unit, a noise amplification unit, and a brightness adjustment unit, which are used to divide the training set T into n independent subsets and preprocess the subsets to generate n course subsets with progressively increasing difficulty gradients.

[0124] The difficulty scheduling module is used to adjust the training course subset in real time according to the training situation. When the training loss value of the previous stage tends to converge, the training course subset of the next stage will be used.

[0125] The ternary feature multi-dimensional loss module includes a feature extraction unit and a multi-dimensional loss function unit. It is used to input the data of the i-th course subset into the feature extraction unit, the ternary branch to obtain the feature vectors of anchor samples, positive samples and negative samples, and the multi-dimensional loss function unit to obtain the vector difference between each group of samples, which is used as the loss value for this iteration for training.

[0126] The feature extraction unit training module is used for backpropagation to update parameters and train the feature extraction unit in the ternary feature multidimensional loss module to obtain the trained feature extraction unit.

[0127] The multi-dimensional measurement module is used in the prediction stage to input the support set S and the query set Q into the trained feature extraction unit to obtain feature vectors of specific dimensions, and to complete image classification for each query and support sample based on the feature vectors of specific dimensions.

[0128] A computer-readable storage medium storing a computer program that, when executed by a processor, implements steps such as those of a method for identifying all types of ships in a disturbed environment based on a progressive curriculum.

[0129] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment, depending on actual needs.

[0130] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0131] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0132] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.

[0133] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.

Claims

1. A method for identifying all types of ships under disturbed environments based on a progressive curriculum, characterized in that, Including the following: Acquire images of the vessel to be identified; The image of the ship to be identified is identified by a pre-trained, progressively-learned, all-type ship recognition model, and the recognition result is obtained. Specifically, the ship type identification model based on a progressive curriculum is trained in the following way: Construct a training set T, a support set S, and a query set Q for ship images; The training set T is divided into n independent subsets, and the subsets are preprocessed to generate n course subsets with progressively increasing difficulty gradients. Difficulty scheduling is performed on the n course subsets. When the training loss value of the previous stage tends to converge, the training course subset of the next stage is used. A ternary feature multidimensional loss module is constructed. The ternary feature multidimensional loss module includes a feature extraction unit and a multidimensional loss function unit. The data of the i-th course subset is input into the feature extraction unit. The feature vectors of anchor samples, positive samples and negative samples are obtained by the ternary branch. The vector difference between each group of samples is obtained by the multidimensional loss function unit and used as the loss value for this iteration for training. Backpropagation updates the parameters, and the feature extraction unit in the ternary feature multidimensional loss module is trained to obtain the trained feature extraction unit. In the prediction phase, the support set S and the query set Q are input into the trained feature extraction unit to obtain feature vectors of a specific dimension. Based on the feature vectors of the specific dimension, image classification is performed on each query and support sample. Specifically, the subset is preprocessed by random cropping, noise amplification, and brightness adjustment to generate n course subsets with progressively increasing difficulty levels. ; Specifically, there are 6 course subsets, i.e., n=6. The preprocessing process is as follows: The initial training curriculum is generated by randomly cropping the data as follows. in, Represents the original image, and Indicates that by application The image obtained by processing the random cropping function. This indicates the size of the image after cropping; By adding noise interference to the initial course as follows, a mid-term training course is generated. in, Indicates the first Line number Liede The pixel values ​​of each channel after adding noise. This represents the probability value of noise addition, used to control the degree of noise interference added to the image; By adjusting the brightness of the mid-term courses as follows, the final training course can be generated. in, This indicates the brightness adjustment value. Indicates the image number 1 Line number Liede The pixel values ​​of each channel; Specifically, the multi-dimensional loss function unit is constructed based on the multi-dimensional triplet fusion loss function, which is expressed as follows: } } The sample is , For mapping functions, ,in It is possible , , If the size of the entire training set is N, then the input (a, p, n) for each triplet is... The Euclidean distance metric between Anchor and Positive samples. This is the Euclidean distance metric between Anchor and Negative samples. The cosine distance metric between Anchor and Positive samples. The cosine distance metric between Anchor and Negative samples; The Minkowski distance is a metric for the distance between Anchor and Positive samples. The Minkowski distance metric between Anchor and Negative samples; The Manhattan distance metric between Anchor and Positive samples. The Manhattan distance metric between Anchor and Negative samples; Different weighting parameters are used to balance the influence between various distance metrics; and The minimum interval between the distances, This indicates that when the value within [ ] is greater than zero, that value is taken as the loss.

2. The method for identifying all types of ships under disturbed environments based on a progressive curriculum, as described in claim 1, is characterized in that... Specifically, difficulty scheduling is performed on the n subsets of courses, including the following: The training course subset used is adjusted in real time based on the training progress. When the training loss value of the previous stage tends to converge, the training course subset for the next stage begins to be used, as shown in the following formula. Where t represents the t-th iteration, and ε represents the convergence threshold of accuracy; The function represents the first time... Accuracy after the next iteration. This indicates the convergence of accuracy calculations starting from this iteration; when the number of iterations exceeds... When, the change in accuracy after each iteration All are less than the preset threshold. This marks the end of this segment of the course.

3. The method for identifying all types of ships under disturbed environments based on a progressive curriculum as described in claim 1, characterized in that, The data from the i-th course subset is input into the feature extraction unit, and the feature vectors of anchor samples, positive samples, and negative samples are obtained using a ternary branch, including: Where n is the index of the course subset, It is a convolutional neural network model. These are the network mapping function and parameters of the convolutional neural network model, respectively. for The extracted feature vector, i takes .

4. The method for identifying all types of ships under disturbed environments based on a progressive curriculum, as described in claim 1, is characterized in that... Specifically, backpropagation updates parameters, including: By summing and subtracting the feature vectors of each branch, and using the sum as the network loss value for backpropagation, we can obtain the gradient information of the input image based on the loss value. , , Using loss function After inverting the gradient of the values, the obtained data are for three ship image samples. , , gradient information, This indicates the differentiation operation.

5. The method for identifying all types of ships under disturbed environments based on a progressive curriculum as described in claim 1, characterized in that, Specifically, the process of completing image classification includes the following: The support set S and query set Q are input into the trained feature extraction unit. The feature vectors of the support set S and query set Q are matched, and the pair with the highest similarity is selected as the predicted category label to complete image classification. The matching is performed using the following formula: in, Different weighting parameters are used to balance the influence between various distance metrics; for The similarity value between samples is used as the confidence score for each class. The category of the group with the highest cosine similarity among all groups is used as the class of that group. Predicted class label for the sample.

6. A device for identifying all types of ships under disturbed environments based on a progressive curriculum, characterized in that: The apparatus comprising the steps of the method according to any one of claims 1-5, wherein the apparatus includes the following: The data acquisition module is capable of acquiring images of the vessel to be identified; The data recognition module is used to recognize the image of the ship to be identified by a pre-trained, progressive course-based ship all-type recognition model, and obtain the recognition result. The pre-trained, progressively structured, ship-wide identification model is obtained through the following modules. The dataset construction module is used to construct the training set T, support set S, and query set Q of ship images. The difficulty generation module includes a random cropping unit, a noise amplification unit, and a brightness adjustment unit, which are used to divide the training set T into n independent subsets and preprocess the subsets to generate n course subsets with progressively increasing difficulty gradients. The difficulty scheduling module is used to adjust the training course subset in real time according to the training situation. When the training loss value of the previous stage tends to converge, the training course subset of the next stage will be used. The ternary feature multi-dimensional loss module includes a feature extraction unit and a multi-dimensional loss function unit. It is used to input the data of the i-th course subset into the feature extraction unit, the ternary branch to obtain the feature vectors of anchor samples, positive samples and negative samples, and the multi-dimensional loss function unit to obtain the vector difference between each group of samples, which is used as the loss value for this iteration for training. The feature extraction unit training module is used for backpropagation to update parameters and train the feature extraction unit in the ternary feature multidimensional loss module to obtain the trained feature extraction unit. The multi-dimensional measurement module is used in the prediction stage to input the support set S and the query set Q into the trained feature extraction unit to obtain feature vectors of specific dimensions, and to complete image classification for each query and support sample based on the feature vectors of specific dimensions.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-5.

Images