Target sonar image recognition submarine search and rescue method based on transfer learning improved VGG-16 model

Abstract

The present invention is applicable to the technical field of ocean engineering. Provided is a target sonar image recognition submarine search and rescue method based on a transfer learning improved VGG-16 model. The method comprises: step one, preprocessing a data set; step two, improving a VGG-16 model; and step three, freezing a transfer policy. In the method, the structure of the VGG-16 model is improved, i.e., on the basis of the characteristics of submarine sonar image data, structural improvement is performed on the VGG-16 model to optimize the network hierarchy and parameter settings, such that the VGG-16 model is more applicable to small-sample target sonar image recognition tasks in a marine environment; the pre-trained improved VGG-16 model is used to perform transfer learning on a target sonar image recognition task for submarine search and rescue, and by means of freezing and adjusting model parameters, the improved VGG-16 model is adaptive to the characteristics of the submarine sonar image data, thereby improving the recognition accuracy; and by means of introducing a data augmentation policy, an effective data augmentation policy is introduced, the diversity of training data is enhanced, and the robustness and generalization capability of the transfer-improved VGG-16 model are improved.

Description

Improved VGG-16 model based on transfer learning for target acoustic image recognition and submarine search and rescue Technical Field

[0001] The present invention relates to the field of marine engineering technology, and more particularly, to a target acoustic image recognition submarine search and rescue method based on an improved VGG-16 model based on transfer learning. Background Art

[0002] Target acoustic pattern recognition is a crucial technology in underwater search and rescue missions. Traditional methods, limited by the complexity and noise of the underwater environment, struggle to accurately identify target acoustic patterns. Recent advances in deep learning technology have opened up new possibilities. The VGG-16 model, a classic deep convolutional neural network, still has certain limitations in underwater search and rescue missions, especially when data is scarce. The VGG-16 model has the following shortcomings:

[0003] (1) Excessive number of parameters: The VGG-16 model has many layers and parameters, which makes its application in submarine search and rescue missions limited by computing resources. Undersea search and rescue missions are usually carried out in resource-constrained environments, so a more lightweight model is needed to achieve target acoustic image recognition.

[0004] (2) Limited feature extraction capabilities: Although VGG-16 performs well in general image recognition tasks, its feature extraction capabilities may be limited in acoustic image recognition tasks in submarine environments. The complexity of the submarine environment and noise interference may cause the features extracted by VGG-16 to be not accurate or discriminative enough, affecting recognition performance.

[0005] (3) Lack of learning ability for specific seabed features: The seabed environment has many unique visual features, such as seabed topography and seabed organisms, which play an important role in the recognition of target acoustic images. However, the VGG-16 model may lack the ability to learn these specific seabed features, resulting in its poor performance in submarine search and rescue missions.

[0006] (4) Insufficient generalization ability: Since the VGG-16 model may not be trained on a rich enough sample set on the submarine search and rescue dataset, its generalization ability may be insufficient. When faced with new submarine environments or target acoustic image types, the VGG-16 model may not adapt well and need to be readjusted and trained, resulting in a decrease in recognition performance.

[0007] (5) Sensitivity to data sparsity: Data acquisition costs for underwater search and rescue missions are high, and labeled data is often insufficient. The VGG-16 model may have difficulty fully utilizing limited labeled data in data-scarce situations, resulting in poor performance. Summary of the Invention

[0008] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a target acoustic image recognition submarine search and rescue method based on transfer learning to improve the VGG-16 model. By improving the VGG-16 model structure, the VGG-16 model is structurally improved according to the characteristics of submarine acoustic image data, and the network hierarchy and parameter settings are optimized to make it more suitable for small sample target acoustic image recognition tasks in marine environments; using the pre-trained improved VGG-16 model, transfer learning is performed on the submarine search and rescue target acoustic image recognition task, and by freezing and adjusting the model parameters, the model is adapted to the characteristics of submarine acoustic image data, thereby improving the recognition accuracy.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] The target acoustic image recognition underwater search and rescue method based on transfer learning to improve the VGG-16 model includes the following steps:

[0011] Step 1: Dataset preprocessing

[0012] Perform a series of operations such as transforming the original data or adding noise to generate more, richer, and more realistic training samples, thereby improving the generalization ability of the model;

[0013] Step 2: Improve the VGG-16 model

[0014] The model structure is simplified based on the traditional VGG16 model, constructing a model consisting of 9 convolutional layers, 5 pooling layers, 2 fully connected layers based on the activation function ReLU, and 9 batch normalization layers placed after each convolutional layer;

[0015] Step 3: Freeze the migration strategy

[0016] A frozen transfer strategy is adopted for the improved VGG-16 model. The pre-trained improved VGG-16 model is used to perform transfer learning on the submarine search and rescue target acoustic image recognition task. By freezing and adjusting the model parameters, it is adapted to the characteristics of the submarine acoustic image data, thereby improving the recognition accuracy of the model.

[0017] The present invention is further configured as follows: the data enhancement strategy introduced in the data set preprocessing in the first step includes adding noise, random rotation, horizontal flipping and manual supervision segmentation.

[0018] The present invention is further configured as follows: the dataset preprocessing in the first step is to make the original dataset more diverse by adding noise, random rotation, horizontal flipping and manual supervision segmentation data enhancement strategy, so that the model is exposed to more diverse data situations during the training process, which facilitates the understanding and generalization of various data patterns.

[0019] The present invention is further configured as follows: the data set preprocessing in the first step is to simulate the noise and deformation in the real world by introducing noise, random rotation and horizontal flipping data enhancement strategies, so that the model gradually learns the robustness to these interference factors during the training process, and the model can more accurately identify the target when facing real data.

[0020] The present invention is further configured as follows: the data set preprocessing in the first step is to increase the diversity of data during the training process by introducing a data enhancement strategy, so that the model learns the general rules of the data, so that the model can make accurate predictions when encountering unseen data.

[0021] The present invention is further configured as follows: the improved VGG-16 model in the second step is to reduce the number of network layers and parameters, reduce computational complexity, accelerate training and real-time inference, alleviate overfitting problems, and reasonably design network layers and activation functions, and introduce batch normalization layers.

[0022] The present invention is further configured as follows: the frozen migration strategy in the third step is to use these feature representations to initialize the model of the submarine search and rescue target acoustic image recognition task through transfer learning, thereby accelerating the convergence process of the model and enabling the model to adapt to the characteristics of the new task more quickly. During the transfer learning process, all 9 convolutional layer parameters of the improved VGG-16 model trained on ImageNet are frozen, the representation ability of the pre-trained model on image features is retained, and only the parameters of the last 3 fully connected layers are adjusted and trained to enable it to better adapt to the characteristics of the submarine acoustic image data.

[0023] The present invention is further configured as follows: the frozen migration strategy described in the third step is to help the model better understand the data of the new task by using the general features learned by the pre-trained model on a large-scale image dataset, and by fine-tuning the parameters of the model to make the model better adapt to the specific data distribution of the new task, thereby improving the recognition accuracy of the model in the underwater search and rescue target acoustic image recognition task.

[0024] The advantages of the present invention are:

[0025] 1. This paper improves the VGG-16 model structure based on the characteristics of seabed acoustic image data, optimizes the network hierarchy and parameter settings, and makes it more suitable for small sample target acoustic image recognition tasks in marine environments.

[0026] 2. The present invention applies a migration strategy and uses a pre-trained improved VGG-16 model to perform transfer learning on the submarine search and rescue target acoustic image recognition task. By freezing and adjusting the model parameters, the model is adapted to the characteristics of the submarine acoustic image data, thereby improving the recognition accuracy.

[0027] 3. The present invention introduces effective data enhancement strategies, such as adding noise, random rotation, horizontal flipping and manual supervision segmentation, which increases the diversity of training data and improves the robustness and generalization ability of the migration-improved VGG-16 model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] FIG1 is a flowchart of the steps of the underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model of transfer learning according to the present invention.

[0029] FIG2 is a submarine aircraft side-scan sonar image dataset of the present invention.

[0030] FIG3 is a side-scan sonar image dataset of a submarine shipwreck according to the present invention.

[0031] FIG4 is a schematic diagram showing the cutting of various parts of the artificially supervised submarine aircraft of the present invention.

[0032] FIG5 is a partial image after image enhancement according to the present invention.

[0033] FIG6 is a schematic diagram of a traditional VGG-16 model of the present invention.

[0034] FIG7 is a schematic diagram of the improved VGG-16 model of FIG5 according to the present invention.

[0035] FIG8 is an improved convolutional neural network model of the present invention.

[0036] FIG9 is a table showing the accuracy and recognition performance of the model of the present invention.

[0037] FIG10 is a VGG-16 P-R curve diagram of the present invention.

[0038] FIG11 is a P-R curve diagram of the improved VGG-16 of the present invention.

[0039] FIG12 is a graph showing the migration-improved VGG-16P-R curve of the present invention.

[0040] FIG13 is a graph showing the training accuracy of the three models of the present invention.

[0041] FIG14 is a graph showing the verification accuracy of the three models of the present invention.

[0042] FIG15 is a graph showing the training loss values ​​of the three models of the present invention.

[0043] FIG16 is a graph showing the validation loss values ​​of the three models of the present invention.

[0044] FIG17 is a VGG-16 confusion matrix table of the present invention.

[0045] FIG18 is a table of the improved VGG-16 confusion matrix of the present invention.

[0046] FIG19 is a table of the confusion matrix of the migration-improved VGG-16 of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by ordinary technicians in the technical field to which this application belongs.

[0049] Example 1

[0050] Please refer to Figures 1-19. The present invention provides the following technical solutions:

[0051] The target acoustic image recognition submarine search and rescue method based on the improved VGG-16 model based on transfer learning specifically includes the following steps:

[0052] Step 1: Dataset Preprocessing

[0053] Perform a series of operations such as transforming the original data or adding noise to generate more, richer, and more realistic training samples, thereby improving the generalization ability of the model.

[0054] Step 2: Improve the VGG-16 model

[0055] The model structure is simplified based on the traditional VGG16 model, constructing a model consisting of 9 convolutional layers, 5 pooling layers, 2 fully connected layers based on the activation function ReLU, and 9 batch normalization layers placed after each convolutional layer.

[0056] Step 3: Freeze the migration strategy

[0057] A frozen transfer strategy is adopted for the improved VGG-16 model. The pre-trained improved VGG-16 model is used to perform transfer learning on the submarine search and rescue target acoustic image recognition task. By freezing and adjusting the model parameters, it is adapted to the characteristics of the submarine acoustic image data, thereby improving the recognition accuracy of the model.

[0058] The working principle of the first embodiment of the present invention is as follows: by improving the structure of the VGG-16 model, the structure of the VGG-16 model is improved according to the characteristics of the seabed acoustic image data, and the network hierarchy and parameter settings are optimized to make it more suitable for the small sample target acoustic image recognition task in the marine environment; by applying the migration strategy, the pre-trained improved VGG-16 model is used to perform transfer learning on the submarine search and rescue target acoustic image recognition task, and by freezing and adjusting the model parameters, the model is adapted to the characteristics of the seabed acoustic image data, thereby improving the recognition accuracy; by introducing effective data enhancement strategies, such as adding noise, random rotation, horizontal flipping and manually supervised segmentation, the diversity of the training data is increased, and the robustness and generalization ability of the migration-improved VGG-16 model are improved.

[0059] Example 2

[0060] Please refer to Figure 1-19. This embodiment 2 makes the following improvements on the basis of embodiment 1. Specifically, the data enhancement strategies introduced in the data set preprocessing in the first step include adding noise, random rotation, horizontal flipping and manually supervised segmentation; the data set preprocessing in the first step is to make the original data set more diversified by adding noise, random rotation, horizontal flipping and manually supervised segmentation data enhancement strategies, so that the model is exposed to more diverse data situations during the training process, which is convenient for understanding and generalizing various data patterns; the data set preprocessing in the first step is to simulate the noise and deformation in the real world by introducing noise, random rotation and horizontal flipping data enhancement strategies, so that the model gradually learns the robustness to these interference factors during the training process, and the model can more accurately identify the target when facing real data; the data set preprocessing in the first step is to increase the diversity of data during the training process by introducing data enhancement strategies, so that the model learns the general laws of the data, so that the model can make accurate predictions when encountering unseen data.

[0061] Working Principle of the Second Embodiment: As shown in Figures 2-3, the original seabed ship and aircraft side-scan sonar datasets are transformed or subjected to a series of operations such as adding noise to the original data to generate more, richer, and more realistic training samples, thereby improving the generalization ability of the model. The data enhancement strategies introduced in this patent include adding noise, random rotation, horizontal flipping, and manually supervised segmentation. Figures 4 and 5 are schematic diagrams of the effects after data enhancement. The introduction of these strategies brings the following innovative effects to the model:

[0062] (1) Increase the diversity of training data: Adding data augmentation strategies such as noise, random rotation, horizontal flipping, and manually supervised segmentation can make the original dataset more diverse. In this way, the model is exposed to more diverse data situations during training, thereby better understanding and generalizing various data patterns, improving the robustness and generalization ability of the model.

[0063] (2) Improving the model's robustness to noise and deformation: Introducing data augmentation strategies such as noise, random rotation, and horizontal flipping can simulate real-world noise and deformation, allowing the model to gradually learn robustness to these interference factors during training. This allows the model to more accurately identify targets when faced with real data, improving its reliability in practical applications.

[0064] (3) Improving the generalization ability of the model: The introduction of data augmentation strategies can not only increase the diversity of training data, but also help alleviate the overfitting problem of the model. By increasing the diversity of data during training, the model is more able to learn the general laws of the data rather than relying too much on specific samples of training data. Therefore, the model can make accurate predictions when encountering unseen data, improving the generalization ability of the model.

[0065] Example 3

[0066] Please refer to Figure 1-19. This third embodiment makes the following improvements based on the second embodiment. Specifically, the VGG-16 model is improved in the second step by reducing the number of network layers and parameters, reducing computational complexity, accelerating training and real-time inference, and alleviating overfitting problems by reasonably designing network layers and activation functions, and introducing batch normalization layers.

[0067] Working principle of the third embodiment: As shown in Figure 6, the traditional VGG-16 model structure has a total of 16 layers, including 9 convolutional layers and 3 fully connected layers. The main components of the network include convolutional layers, maximum pooling layers, and fully connected layers. In order to adapt to the characteristics of the small number of seabed target acoustic image data samples, while taking into account the factors of training accuracy and training efficiency, improvements were made on the basis of the traditional VGG16 model. First, the model structure was simplified to construct a model consisting of 9 convolutional layers, 5 pooling layers, 2 fully connected layers based on the activation function ReLU, and 9 batch normalization layers placed after each convolutional layer. Figures 7 and 8 are schematic diagrams of the improved VGG-16 model. The improved VGG-16 model brings the following effects to the innovation point:

[0068] (1) Reduced computational complexity: By reducing the number of layers and parameters in the network, the simplified model has lower computational complexity. This means that the model requires fewer computing resources during the training and inference phases, enabling faster training and real-time inference, thus improving model efficiency.

[0069] (2) Alleviate overfitting: The simplified model structure reduces the number of parameters and the complexity of the model, which helps to alleviate the overfitting problem. Overfitting is a common problem in small-sample target acoustic image recognition tasks, and the simplified model structure can better adapt to limited training data and improve the generalization ability of the model.

[0070] (3) Improved training speed: By reducing the number of model parameters and computational complexity, the simplified model can complete the training process faster. This means that the model training process can be iterated faster, accelerating the model convergence speed, and thus achieving better model performance faster.

[0071] (4) Maintaining high performance: Despite the simplified model structure, the simplified model can still maintain a high level of performance by rationally designing the network layers and activation functions, as well as introducing batch normalization layers. This means that the innovative VGG-16 model can achieve comparable or even better recognition results than the original VGG-16 model in small-sample target sound image recognition tasks.

[0072] Example 4

[0073] Please refer to Figure 1-19. This fourth embodiment makes the following improvements on the basis of the third embodiment. Specifically, in the third step, the frozen migration strategy is to use these feature representations to initialize the model of the submarine search and rescue target acoustic image recognition task through transfer learning, thereby accelerating the convergence process of the model and enabling the model to adapt to the characteristics of the new task more quickly. In the transfer learning process, all 9 convolutional layer parameters of the improved VGG-16 model trained on ImageNet are frozen, and the representation ability of the pre-trained model on image features is retained. Only the parameters of the last 3 fully connected layers are adjusted and trained so that it can better adapt to the characteristics of the submarine acoustic image data; in the third step, the frozen migration strategy is to help the model better understand the data of the new task by using the general features learned by the pre-trained model on a large-scale image dataset, and by fine-tuning the parameters of the model, the model is better adapted to the specific data distribution of the new task, thereby improving the recognition accuracy of the model in the submarine search and rescue target acoustic image recognition task.

[0074] The working principle of this fourth embodiment is to adopt a frozen transfer strategy for the improved VGG-16 model. Using the pre-trained improved VGG-16 model, transfer learning is performed on the underwater search and rescue target acoustic image recognition task. By freezing and adjusting the model parameters, it adapts to the characteristics of the underwater acoustic image data, thereby improving the model's recognition accuracy. The frozen improved VGG-16 model brings the following innovative effects:

[0075] (1) Leveraging the features of pre-trained models: The pre-trained improved VGG-16 model has learned rich feature representations on a large-scale image dataset. Through transfer learning, these feature representations are used to initialize the model for the underwater search and rescue target acoustic image recognition task, thereby accelerating the model's convergence process and enabling the model to adapt to the characteristics of new tasks more quickly.

[0076] (2) Freezing and adjusting model parameters: During the transfer learning process, all nine convolutional layer parameters of the modified VGG-16 model trained on ImageNet were frozen to preserve the pre-trained model’s ability to represent image features. Only the parameters of the last three fully connected layers were adjusted and trained to better adapt them to the characteristics of the seabed acoustic image data. This freezing strategy not only leverages the general features learned by the pre-trained model but also fine-tunes the model parameters to adapt it to the specific characteristics of the seabed acoustic image data, thereby improving recognition accuracy.

[0077] (3) Improving the generalization ability of the model: Using pre-trained models for transfer learning can improve the generalization ability of the model. The general features learned by the pre-trained model on large-scale image datasets can help the model better understand the data of the new task, and by fine-tuning the model parameters, the model can be better adapted to the specific data distribution of the new task, thereby improving the recognition accuracy of the model in the underwater search and rescue target acoustic image recognition task.

[0078] As shown in Figure 9, the transfer-improved VGG-16 model achieved a training accuracy of 97.00% on the training set, a 2.20% and 0.88% improvement over the traditional VGG-16 and improved VGG-16 models, respectively. The validation set accuracy also reached 97.70%, exceeding the other two methods by 9.50% and 6.4%, respectively. This demonstrates the feasibility of the transfer-improved VGG-16 model. The frozen convolutional layers are fully trained and effectively extract common image features. The initialized and retrained high-level parameters effectively recognize abstract features in side-scan sonar target images, achieving excellent recognition accuracy. The transfer-improved VGG-16 model achieved a loss of 0.08 on both the training and validation sets, significantly lower than the traditional VGG-16 and improved VGG-16 models, demonstrating that the transfer-improved VGG-16 model has a better fit and is more universal.

[0079] Figures 10-12 show the PR curves for the three models. Among the AP values ​​of the transfer-improved VGG-16 model, the traditional VGG-16 model, and the improved VGG-16 model, the transfer-improved VGG-16 model achieved the highest AP, 97.14%, 7.28% and 6.09% higher than the traditional VGG-16 and improved VGG-16 models, respectively. As shown in the PR curves for the three models in Figures 10-12, the transfer-improved VGG-16 model achieved higher recall and precision than the traditional VGG-16 and improved VGG-16 models. All indicators demonstrate that the transfer-improved VGG-16 model outperformed the traditional VGG-16 and improved VGG-16 models in recognition performance.

[0080] To better reflect the real-time changes and fit of model training and validation, the model evaluation results are saved after each training step. The loss values ​​and accuracy rates of the training and test sets of the three models are shown in Figures 13-15.

[0081] As shown in Figures 13-14, the initial training accuracies of the transfer-improved VGG-16 model, the improved VGG-16 model, and the traditional VGG-16 model were 58.36%, 54.02%, and 28.08%, respectively. Their initial validation accuracies were 59.50%, 58.10%, and 29.90%, respectively. The transfer-improved VGG-16 model achieved higher recognition accuracy than the other two models during both the initial training and validation phases. The transfer-improved VGG-16 model achieved higher recognition accuracy than the other two models in every round of training on the training set, reaching 91% after 21 steps and converging to approximately 96% after 57 steps. The traditional VGG-16 model, on the other hand, only reached stability after 75 steps, achieving a recognition accuracy of approximately 93%. The transfer-improved VGG-16 model reached convergence after 53 training steps on the validation set, achieving a recognition accuracy of approximately 97%. Although the recognition accuracy fluctuated significantly in the first 20 validation rounds, it stabilized after 20 rounds and was significantly higher than the other two models. As shown in Figures 15-16, the transfer-improved VGG-16 model achieved lower training and validation loss than the other two models, and converged earlier in the validation rounds.

[0082] Figures 17-19 are the confusion matrices corresponding to the validation sets of the three models: the migration-improved VGG-16 model, the improved VGG-16 model, and the traditional VGG-16 model. In the confusion matrix corresponding to the validation set of the traditional VGG-16 model, 60 images with the label "seabed" were misidentified as wings, 4 images with the label "tail" were misidentified as wings; 7 images with the label "wing" were misidentified as seabed, 1 image with the label "tail" and 1 image with the label "shipwreck" were misidentified as shipwreck; 4 images with the label "airplane" were misidentified as shipwreck; 1 image with the label "shipwreck" was misidentified as seabed, 1 image with the label "tail", 2 images with the label "wing" were misidentified as shipwreck.

[0083] In the confusion matrix corresponding to the validation set of the improved VGG-16 model, there is 1 misidentified image as the tail of the aircraft, 32 misidentified images as wings, and 1 misidentified image as a shipwreck when the label is the seabed; there are 2 misidentified images as the seabed and 6 misidentified images as wings when the label is the tail; there are 14 misidentified images as the seabed, 2 misidentified images as the tail, and 1 misidentified image as a shipwreck when the label is the wing; there are 5 misidentified images as the seabed, 2 misidentified images as wings, and 13 misidentified images as aircraft when the label is the shipwreck.

[0084] In the validation set confusion matrix for the transfer-improved VGG-16 model, one image labeled "seabed" was misidentified as an aircraft tail, and two images were misidentified as wings; three images labeled "tail" were misidentified as seabed, and four images were misidentified as wings; three images labeled "wing" were misidentified as seabed; one image labeled "airplane" was misidentified as a shipwreck; and 19 images labeled "shipwreck" were misidentified as aircraft. In summary, the transfer-improved VGG-16 model achieved the lowest recognition error rates for all five types of seabed objects compared to the other two models. Among the five types of seabed objects, the transfer-improved VGG-16 model achieved the highest recognition accuracy for aircraft on the seabed and the lowest for shipwrecks.

[0085] Obviously, the embodiments described above are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0086] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, tasks, devices, components and / or combinations thereof.

[0087] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

[0088] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A target acoustic image recognition method for submarine search and rescue based on the improved VGG-16 model based on transfer learning, characterized by: The following steps are involved: Step 1: Dataset preprocessing Perform a series of operations such as transforming the original data or adding noise to generate more, richer, and more realistic training samples, thereby improving the generalization ability of the model; Step 2: Improve the VGG-16 model The model structure is simplified based on the traditional VGG16 model, constructing a model consisting of 9 convolutional layers, 5 pooling layers, 2 fully connected layers based on the activation function ReLU, and 9 batch normalization layers placed after each convolutional layer; Step 3: Freeze the migration strategy A frozen transfer strategy is adopted for the improved VGG-16 model. The pre-trained improved VGG-16 model is used to perform transfer learning on the submarine search and rescue target acoustic image recognition task. By freezing and adjusting the model parameters, it is adapted to the characteristics of the submarine acoustic image data, thereby improving the recognition accuracy of the model.

2. The underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model through transfer learning according to claim 1 is characterized in that: The data augmentation strategies introduced in the dataset preprocessing described in the first step include adding noise, random rotation, horizontal flipping and manual supervision segmentation.

3. The underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model of transfer learning according to claim 2 is characterized in that: The dataset preprocessing described in the first step is to make the original dataset more diverse by adding noise, random rotation, horizontal flipping and manually supervised segmentation data enhancement strategies. The model is exposed to more diverse data situations during training, which facilitates the understanding and generalization of various data patterns.

4. The underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model through transfer learning according to claim 3 is characterized in that: The dataset preprocessing described in the first step is to simulate the noise and deformation in the real world by introducing noise, random rotation and horizontal flipping data enhancement strategies, so that the model gradually learns the robustness to these interference factors during the training process, and the model can more accurately identify targets when faced with real data.

5. The underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model through transfer learning according to claim 4 is characterized in that: The dataset preprocessing described in the first step is to increase the diversity of data during the training process by introducing a data augmentation strategy, so that the model can learn the general rules of the data and make accurate predictions when encountering unseen data.

6. The underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model through transfer learning according to claim 5, characterized in that: The improvement of the VGG-16 model described in the second step is to reduce the number of network layers and parameters, reduce computational complexity, accelerate training and real-time inference, alleviate overfitting problems, and reasonably design network layers and activation functions, as well as introduce batch normalization layers.

7. The underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model through transfer learning according to claim 6, characterized in that: The frozen migration strategy described in the third step is to use these feature representations to initialize the model of the submarine search and rescue target acoustic image recognition task through transfer learning, thereby accelerating the convergence process of the model and enabling the model to adapt to the characteristics of the new task more quickly. In the process of transfer learning, all 9 convolutional layer parameters of the improved VGG-16 model trained on ImageNet are frozen, retaining the representation ability of the pre-trained model on image features, and only adjusting the parameters of the last 3 fully connected layers and training them to enable them to better adapt to the characteristics of the submarine acoustic image data.

8. The underwater search and rescue method for target acoustic image recognition based on the improved VGG-16 model through transfer learning according to claim 7, characterized in that: The frozen migration strategy described in the third step is to help the model better understand the data of the new task by using the general features learned by the pre-trained model on a large-scale image dataset, and to make the model better adapt to the specific data distribution of the new task by fine-tuning the model parameters, thereby improving the recognition accuracy of the model in the underwater search and rescue target acoustic image recognition task.

Images