Ship category identification method based on incremental learning

Abstract

The invention discloses a ship category identification method based on incremental learning, and the method comprises the steps: carrying out the improvement of a training mechanism and a learning strategy of an original YOLO11 model, inputting a training set 1 into the original YOLO11 model for training, and obtaining the optimal weight of the model; the optimal weight of the model is loaded to the improved YOLO11 model, and an incremental target detection model is constructed through an EWC algorithm and a periodic data playback strategy; training an incremental target detection model by using the training set 2, and introducing a differential regularization strategy to enable the model to obtain the ability of inhibiting disastrous forgetting; and testing the incremental target detection model by using the test set, and continuously adjusting the regularization strength to obtain an optimal detection result. Incremental learning and target detection are combined, new ship categories are continuously learned through the model, meanwhile, stable detection performance of original ship categories is kept, and the problem of knowledge forgetting in the model iteration process is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a ship category recognition method based on incremental learning. Background Art

[0002] Improving the intelligence level of surface vessels is of great significance to maritime safety, shipping management, environmental protection, and other fields. By accurately identifying different types of vessels, shipping scheduling can be optimized, maritime accidents can be prevented, illegal activities can be effectively combated, and the development and management of marine resources can be enhanced. However, traditional object detection models are unable to dynamically expand recognition categories. When faced with new ship types, they often need to integrate datasets from all categories for retraining, which is computationally expensive. If only the newly added ship categories are trained, catastrophic forgetting often occurs. That is, while learning new categories, the recognition ability of the original categories decreases. Summary of the Invention

[0003] The purpose of the present invention is to consider the catastrophic forgetting challenges faced in the incremental learning scenario of ship target detection in a dynamic ocean environment, as well as the high requirements for real-time performance and stability when continuously introducing new categories. A ship category recognition method based on incremental learning is provided. Based on the incremental learning mode, You Only Look Once Version 11 (YOLO11) is improved, and the elastic weight consolidation (EWC) method is combined to fix the weights of important parameters by calculating the importance of the old class model parameters, ensuring that the detection accuracy of new ship classes is improved while maintaining stable recognition of old ship classes. By effectively retaining and updating historical category information, it is ensured that the recognition ability of old categories is not lost while learning new categories, and significantly improving the detection stability, accuracy and cross-scenario generalization performance of the model during the incremental update process. This method not only enhances the adaptability of ship classification tasks, but also provides technical support for the practical application of intelligent marine equipment.

[0004] To achieve the above objectives, the present invention provides a technical solution: a ship category recognition method based on incremental learning, which accurately recognizes ship categories based on an improved YOLO11 model. The YOLO11 model is improved from two aspects: the training mechanism and the learning strategy. In terms of the training mechanism, the elastic weight fixation method (EWC algorithm) is integrated into the trainer module of the YOLO11 model. The EWC algorithm retains old task knowledge by constraining the update amplitude of key parameters of the model. In terms of the learning strategy, a periodic data replay strategy is adopted, that is, m rounds of old task data review are inserted after each predetermined number of new task training rounds, thereby realizing the alternating reinforcement of new and old knowledge.

[0005] The specific implementation of the ship type identification method includes the following steps:

[0006] S1: Obtain a ship image dataset containing data of various ship categories. Use standardized preprocessing to construct an expanded dataset. This dataset is used as a test set and divided into two training sets: training set 1 containing images of several ship categories and training set 2 containing images of the remaining ship categories.

[0007] S2: Input training set 1 into the original YOLO11 model for training to obtain the optimal model weights;

[0008] S3: For the ship detection task, the optimal model weights obtained in step S2 are loaded into the improved YOLO11 model. The EWC algorithm is used to impose elastic constraints on the model parameters. At the same time, a periodic data playback strategy is adopted to achieve alternating reinforcement of new and old knowledge, so that the improved YOLO11 model can obtain incremental learning capabilities. The improved YOLO11 model with incremental learning capabilities is called the incremental target detection model.

[0009] S4: Training set 2 is input into the incremental object detection model for training. At the same time, a differentiated regularization strategy is introduced. When updating the model parameters, large constraints are imposed on the parameters corresponding to the knowledge obtained in training set 1 to prevent them from deviating from the original optimal state. On the other hand, small constraints are imposed on the parameters related to the features of the new samples in training set 2 to ensure that the model has sufficient flexibility to learn and adapt to the newly added targets. This enables the model to have the ability to adaptively learn the features of new tasks and effectively suppress catastrophic forgetting while retaining the detection accuracy of the old tasks.

[0010] S5: Use the test set to test the incremental object detection model trained in step S4. By observing the performance of the sample forgetting rate indicator and the mAP50 key performance indicator, continuously adjust the regularization strength until both indicators reach the ideal state. Then use the regularization strength at this time to fully test the test set and finally obtain the optimal detection result.

[0011] Furthermore, in step S1, the ship image dataset is subjected to standardization preprocessing, including normalizing pixel values ​​to the range of [0, 1], unifying sequence lengths through dynamic padding or truncation, and ensuring dimensional consistency of batch input tensors using shape adjustment techniques.

[0012] Furthermore, in step S2, for the input training set 1, the training operation steps are as follows:

[0013] S21: For the input training set 1, the backbone network of the YOLO11 model is used to extract basic visual features at different levels to capture low-level and high-level information of the image;

[0014] S22: A feature pyramid network is used to fuse basic visual features from different levels, integrating multi-scale information to improve the model's detection ability for objects of different sizes.

[0015] S23: Through multiple iterative training, the loss function is minimized and the model parameters are optimized, and finally the optimal model weight corresponding to the task is obtained.

[0016] Furthermore, in step S4, a differentiated regularization strategy is introduced to achieve the best accuracy of training effect in order to achieve the balance between learning and forgetting of new and old tasks. The specific situation is as follows:

[0017] The training set 2 is input into the incremental target detection model for training. During the training process, in order to achieve the minimum forgetting of the training set 1 and the optimal learning task balance of the training set 2 data, the idea of ​​differential regularization is adopted. That is, when the model parameters are updated, large constraints are imposed on the parameters corresponding to the knowledge obtained in the training set 1 to prevent them from deviating from the original optimal state, while small constraints are imposed on the parameters related to the new sample features in the training set 2 to ensure that the model has sufficient flexibility to learn and adapt to the newly added targets. This enables the model to have the ability to adaptively learn new task features and effectively suppress catastrophic forgetting while retaining the detection accuracy of the old task; different regularization strengths are applied to the parameters of the new and old tasks, and the EWC loss function Loss of the differential regularization strategy is designed. EWC , as follows:

[0018] The Fisher information matrix evaluation method is used to calculate the importance of each parameter in the new and old tasks respectively. This is used to determine which parameters have a greater impact on the original task and which parameters are more critical to the new task. This provides a basis for setting different regularization strengths in the subsequent training process, so that the model can more reasonably balance the retention of old knowledge and the learning of new knowledge during training.

[0019] When calculating the EWC loss function Loss EWC In the example above, a high regularization strength λ is used for the old tasks. old , while for new tasks a low regularization strength λ is used new ,at this time Among them S new 、S old Represent the datasets in the new and old tasks respectively, F i 、F j are the elements of the Fisher information matrix, respectively, which measure the importance of the old task-related parameters indexed as i and the new task-related parameters indexed as j, θ i,new represents the current value of the old task related parameters when learning the new task, θ i,old represents the old task parameter value, θ j,new Indicates the current value of the parameters related to the new task, θj,old Indicates the initial value of the new task parameter;

[0020] In the incremental target detection model, its total loss function Loss total =λ1(αLoss box +βLoss obj +γLoss cls )+λ2LOSS ewc (θ), where Loss box 、Loss obj 、Loss cls Corresponding to bounding box loss, confidence loss and classification loss respectively; LOSS ewc (θ) represents the loss function in the EWC algorithm, λ1, λ2, α, β, and γ are weight coefficients used to balance the contribution of each part of the loss;

[0021] Finally, a large number of comparative experiments are used to dynamically adjust the regularization strength of new and old tasks to achieve the best trade-off between tasks.

[0022] Furthermore, in step S5, the test set is used for testing, and the regularization strength is continuously adjusted, including the following steps:

[0023] S51: Analyze the impact of regularization strength on the performance of the incremental object detection model by monitoring the sample forgetting rate metric and the mAP50 key performance indicator. Based on these two indicators, gradually adjust the regularization strength to achieve the optimal balance between forgetting and learning.

[0024] S52: After adjusting the regularization strength, retest and evaluate the incremental target detection model, and continue to optimize until the incremental target detection model achieves global minimization of the loss function on the training set, ensuring the optimal performance of the incremental target detection model in all indicators.

[0025] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0026] 1. By constraining the update of important parameters, catastrophic forgetting can be prevented when adapting to new tasks (such as new ship types or environmental changes), and the original ship detection capabilities can be maintained. This is especially important for scenarios that require continuous expansion of functions.

[0027] 2. It can adapt to new tasks without old data. Compared with the traditional retraining method that requires the use of both new and old data, it can significantly save storage and computing resources. When the ship detection model is deployed in a new sea area or with a new sensor, there is no need to collect historical data again.

[0028] 3. Suppress drastic parameter fluctuations, improve the stability of the model during incremental learning, ensure stable detection performance, and avoid false detection or missed detection due to model updates in scenarios with high reliability requirements such as ship monitoring.

[0029] 4. Dynamically adjust constraints based on parameter importance, optimize parameters in a targeted manner, protect only key weights, allow non-critical parameters to adapt to new data, and improve parameter utilization efficiency for ship inspection.

[0030] 5. It can gradually adapt to different lighting, weather or sea conditions without overfitting new data, enhance cross-environment generalization capabilities, absorb new environmental specificities while retaining general ship characteristics, and improve robustness in changing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flowchart of the method of the present invention.

[0032] Figure 2 This is the detection result diagram of ship category identification after incremental learning of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0034] like Figure 1 As shown, this embodiment discloses a ship category recognition method based on incremental learning. The method is based on the improved YOLO11 model to achieve accurate recognition of ship categories. The YOLO11 model is improved from two aspects: training mechanism and learning strategy: in terms of training mechanism, the elastic weight fixing method is integrated in the trainer module of the YOLO11 model, referred to as the EWC algorithm. This algorithm retains old task knowledge by constraining the update amplitude of key parameters of the model; in terms of learning strategy, a periodic data playback strategy is adopted, that is, after each predetermined number of new task training rounds, m rounds of old task data review are inserted, so as to achieve alternating reinforcement of new and old knowledge. In the present invention, with 10 epochs as a cycle, it is set to insert 1 round of old task data review after every 9 rounds of new task training; the specific situation is as follows:

[0035] S1: Obtain a ship image dataset containing data of various ship categories, and use standardized preprocessing to construct an expanded dataset. On the one hand, this dataset is used as a test set, and on the other hand, it is divided into two training sets, namely training set 1 containing images of several ship categories and training set 2 containing images of the remaining ship categories;

[0036] In this embodiment, the SeaShips dataset is used for training, where the dataset D = {p1, p2, ..., p j ,...,p m}, contains m images, where the jth image Represented in the set of real numbers It has height H, width W and corresponding RGB3 channels. The jth image p j There is a corresponding annotation b j , the annotation contains five parameters b j =(k j ,u j ,v j ,l j ,t j ), corresponding to the category label k j And the horizontal and vertical coordinate position, width, and height of the prediction box, where k j ∈{e1,e2,e3,e4,e5,e6} represents six different ship categories. Standardization preprocessing is performed on the six types of ship data, including normalizing pixel values ​​to the range of [0,1], unifying sequence lengths through dynamic padding or truncation, and using shape adjustment techniques to ensure dimensional consistency of batch input tensors.

[0037] Table 1 shows the distribution of data in each category:

[0038]

[0039] S2: Input training set 1 into the original YOLO11 model for training to obtain the optimal model weights. The specific steps include the following:

[0040] S21: For the input training set 1, the backbone network of the YOLO11 model is used to extract basic visual features at different levels to capture low-level and high-level information of the image;

[0041] S22: A feature pyramid network is used to fuse basic visual features from different levels, integrating multi-scale information to improve the model's detection ability for objects of different sizes.

[0042] S23: Through multiple iterative training, the loss function is minimized and the model parameters are optimized, and finally the optimal model weight corresponding to the task is obtained.

[0043] S3: For the ship detection task objective, the optimal model weights obtained in step S2 are loaded into the improved YOLO11 model. The EWC algorithm is used to impose elastic constraints on the network parameters. At the same time, a periodic data replay strategy is adopted to achieve alternating reinforcement of new and old knowledge, so that the improved YOLO11 model can obtain incremental learning capabilities. The improved YOLO11 model with incremental learning capabilities is called the incremental target detection model.

[0044] S4: Input training set 2 into the incremental target detection model for training. At the same time, a differentiated regularization strategy is introduced. When updating model parameters, larger constraints are imposed on the parameters corresponding to the knowledge acquired in training set 1 to prevent them from deviating from their original optimal state. On the other hand, smaller constraints are imposed on the parameters related to the features of new samples in training set 2 to ensure that the model has sufficient flexibility to learn and adapt to the newly added targets. This enables the model to adaptively learn the features of new tasks while retaining the detection accuracy of the old tasks and effectively suppressing catastrophic forgetting. Specifically, the steps include:

[0045] Training set 2 is input into the incremental target detection model for training. During the training process, in order to achieve the minimum forgetting of training set 1 and the best learning task balance of training set 2 data, the idea of ​​differential regularization is adopted, that is, when the model parameters are updated, larger constraints are imposed on the parameters corresponding to the knowledge obtained in training set 1 to prevent them from deviating from the original optimal state; and smaller constraints are imposed on the parameters related to the new sample features in training set 2 to ensure that the model has sufficient flexibility to learn and adapt to the new targets. This enables the model to have the ability to adaptively learn new task features and effectively suppress catastrophic forgetting while retaining the detection accuracy of the old task. Different regularization strengths are applied to the parameters of the new and old tasks, and the EWC loss function Loss of the differential regularization strategy is designed. EWC , as follows:

[0046] The Fisher information matrix evaluation method is used to calculate the importance of each parameter in the new and old tasks respectively. This is used to determine which parameters have a greater impact on the original task and which parameters are more critical to the new task. This provides a basis for setting different regularization strengths in the subsequent training process, so that the model can more reasonably balance the retention of old knowledge and the learning of new knowledge during the training process.

[0047] When calculating the EWC loss function Loss EWC In the example above, a high regularization strength λ is used for the old tasks. old , while for new tasks a low regularization strength λ is used new ,at this time Among them S new 、S old Represent the datasets in the new and old tasks respectively, F i 、F j are the elements of the Fisher information matrix, respectively, which measure the importance of the old task-related parameters indexed as i and the new task-related parameters indexed as j, θ i,new represents the current value of the old task related parameters when learning the new task, θ i,old represents the old task parameter value, θ j,new Indicates the current value of the parameters related to the new task, θ j,old Represents the initial value of the new task parameter. In this invention, after experimental verification, when The effect is best when

[0048] In the incremental target detection model, its total loss function Loss total =λ1(αLoss box +βLoss obj +γLoss cls )+λ2LOSS ewc (θ), where Loss box 、Loss obj 、Loss cls Corresponding to bounding box loss, confidence loss and classification loss respectively; LOSS ewc (θ) represents the loss function in the EWC algorithm, λ1, λ2, α, β, and γ are weight coefficients used to balance the contribution of each part of the loss. In the present invention, λ1 = α = β = γ = 1, and λ2 = 0.5 are set;

[0049] Finally, a large number of comparative experiments are used to dynamically adjust the regularization strength of new and old tasks to achieve the best trade-off between tasks.

[0050] Table 2 shows the impact of different weight parameters on the experimental results:

[0051] serial number Weight parameter Old_data(mAP50) New_data(mAP50) All(mAP50) 00 <![CDATA[λ1=1,λ2=1]]> 81.30％ 68.35％ 77.00％ 01 <![CDATA[λ1=0.5,λ2=0.5]]> 81.33％ 64.65％ 75.80％ 02 <![CDATA[λ1=1,λ2=0.5]]> 81.33％ 71.40％ 78.00％ 03 <![CDATA[λ1=0.5,λ2=1]]> 81.88％ 43.45％ 69.00％

[0052] S5: Use the test set to test the incremental object detection model obtained in step S4. By observing the sample forgetting rate and mAP50 key performance indicators, the regularization strength is continuously adjusted until both indicators reach the ideal state. Then, the regularization strength at this time is used to conduct a comprehensive test on the test set, and finally the optimal detection result is obtained, such as Figure 2 shown.

[0053] S51: Analyze the impact of regularization strength on the performance of the incremental object detection model by monitoring the sample forgetting rate and mAP50 key performance indicators. Based on these two indicators, gradually adjust the regularization strength to achieve the optimal balance between forgetting and learning.

[0054] S52: After adjusting the regularization strength, retest and evaluate the incremental target detection model, and continue to optimize until the incremental target detection model achieves global minimization of the loss function on the training set, ensuring the optimal performance of the incremental target detection model in all indicators.

[0055] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A ship category recognition method based on incremental learning, characterized in that: This method is based on the improved YOLO11 model to achieve accurate ship classification. The YOLO11 model is improved from two aspects: training mechanism and learning strategy. In terms of training mechanism, the elastic weight fixation method (EWC algorithm) is integrated into the trainer module of the YOLO11 model. This algorithm retains old task knowledge by constraining the update amplitude of key model parameters. In terms of learning strategy, a periodic data replay strategy is adopted. That is, after every predetermined number of rounds of new task training, m rounds of old task data review are inserted to achieve alternating reinforcement of new and old knowledge. The specific implementation of the ship type identification method includes the following steps: S1: Obtain a ship image dataset containing data of various ship categories. Use standardized preprocessing to construct an expanded dataset. This dataset is used as a test set and divided into two training sets: training set 1 containing images of several ship categories and training set 2 containing images of the remaining ship categories. S2: Input training set 1 into the original YOLO11 model for training to obtain the optimal model weights; S3: For the ship detection task, the optimal model weights obtained in step S2 are loaded into the improved YOLO11 model. The EWC algorithm is used to impose elastic constraints on the model parameters. At the same time, a periodic data playback strategy is adopted to achieve alternating reinforcement of new and old knowledge, so that the improved YOLO11 model can obtain incremental learning capabilities. The improved YOLO11 model with incremental learning capabilities is called the incremental target detection model. S4: Training set 2 is input into the incremental object detection model for training. At the same time, a differentiated regularization strategy is introduced. When updating the model parameters, large constraints are imposed on the parameters corresponding to the knowledge obtained in training set 1 to prevent them from deviating from the original optimal state. On the other hand, small constraints are imposed on the parameters related to the features of the new samples in training set 2 to ensure that the model has sufficient flexibility to learn and adapt to the newly added targets. This enables the model to have the ability to adaptively learn the features of new tasks and effectively suppress catastrophic forgetting while retaining the detection accuracy of the old tasks. S5: Use the test set to test the incremental object detection model trained in step S4. By observing the performance of the sample forgetting rate indicator and the mAP50 key performance indicator, continuously adjust the regularization strength until both indicators reach the ideal state. Then use the regularization strength at this time to fully test the test set and finally obtain the optimal detection result.

2. The ship classification method based on incremental learning according to claim 1, characterized in that: In step S1, the ship image dataset is preprocessed by standardization, including normalizing pixel values ​​to the range of [0, 1], unifying sequence lengths through dynamic padding or truncation, and ensuring the dimensional consistency of batch input tensors using shape adjustment techniques.

3. The ship classification method based on incremental learning according to claim 2, characterized in that: In step S2, for the input training set 1, the training operation steps are as follows: S21: For the input training set 1, the backbone network of the YOLO11 model is used to extract basic visual features at different levels to capture low-level and high-level information of the image; S22: A feature pyramid network is used to fuse basic visual features from different levels, integrating multi-scale information to improve the model's detection ability for objects of different sizes. S23: Through multiple iterative training, the loss function is minimized and the model parameters are optimized, and finally the optimal model weight corresponding to the task is obtained.

4. The ship classification method based on incremental learning according to claim 3, characterized in that: In step S4, a differentiated regularization strategy is introduced to achieve the best training accuracy in order to achieve a balance between learning and forgetting of new and old tasks. The specific situation is as follows: The training set 2 is input into the incremental target detection model for training. During the training process, in order to achieve the minimum forgetting of the training set 1 and the optimal learning task balance of the training set 2 data, the idea of ​​differential regularization is adopted. That is, when the model parameters are updated, large constraints are imposed on the parameters corresponding to the knowledge obtained in the training set 1 to prevent them from deviating from the original optimal state, while small constraints are imposed on the parameters related to the new sample features in the training set 2 to ensure that the model has sufficient flexibility to learn and adapt to the newly added targets. This enables the model to have the ability to adaptively learn new task features and effectively suppress catastrophic forgetting while retaining the detection accuracy of the old task; different regularization strengths are applied to the parameters of the new and old tasks, and the EWC loss function Loss of the differential regularization strategy is designed. EWC , as follows: The Fisher information matrix evaluation method is used to calculate the importance of each parameter in the new and old tasks respectively. This is used to determine which parameters have a greater impact on the original task and which parameters are more critical to the new task. This provides a basis for setting different regularization strengths in the subsequent training process, so that the model can more reasonably balance the retention of old knowledge and the learning of new knowledge during training. When calculating the EWC loss function Loss EWC In the example above, a high regularization strength λ is used for the old tasks. old , while for new tasks a low regularization strength λ is used new ,at this time Among them S new 、S old Represent the datasets in the new and old tasks respectively, F i 、F j are the elements of the Fisher information matrix, respectively, which measure the importance of the old task-related parameters indexed as i and the new task-related parameters indexed as j, θ i,new represents the current value of the old task related parameters when learning the new task, θ i,old represents the old task parameter value, θ j,new Indicates the current value of the parameters related to the new task, θ j,old Indicates the initial value of the new task parameter; In the incremental target detection model, its total loss function Loss total =λ1(αLoss box +βLoss obj +γLoss cls )+λ2LOSS ewc (θ), where Loss box 、Loss obj 、Loss cls Corresponding to bounding box loss, confidence loss and classification loss respectively; LOSS ewc (θ) represents the loss function in the EWC algorithm, λ1, λ2, α, β, and γ are weight coefficients used to balance the contribution of each part of the loss; Finally, a large number of comparative experiments are used to dynamically adjust the regularization strength of new and old tasks to achieve the best trade-off between tasks.

5. The ship category identification method based on incremental learning according to claim 4, characterized in that: In step S5, the test set is used for testing, and the regularization strength is continuously adjusted, including the following steps: S51: Analyze the impact of regularization strength on the performance of the incremental object detection model by monitoring the sample forgetting rate metric and the mAP50 key performance indicator. Based on these two indicators, gradually adjust the regularization strength to achieve the optimal balance between forgetting and learning. S52: After adjusting the regularization strength, retest and evaluate the incremental target detection model, and continue to optimize until the incremental target detection model achieves global minimization of the loss function on the training set, ensuring the optimal performance of the incremental target detection model in all indicators.

Images