Maritime unmanned platform target recognition method based on knowledge distillation and semi-supervised learning

Abstract

The application discloses a maritime unmanned platform target identification method based on knowledge distillation and semi-supervised learning. Firstly, in order to improve the identification ability of the model and the semi-supervised learning ability, color space transformation (such as brightness adjustment and contrast enhancement), rank one prior image enhancement (ROP image enhancement) and random masking enhancement techniques are established, and the combination of multiple data enhancement techniques improves the identification ability of the model under complex lighting and shielding conditions. Secondly, through semi-supervised learning, a small amount of labeled image data and a large amount of unlabeled image data are effectively combined for training, and under the condition of a small amount of labeled image data, the model identification ability is close to supervised learning. Finally, the knowledge distillation technology is introduced, the knowledge of the teacher model is effectively transmitted, the identification ability of the student model is significantly improved, and the parameter quantity is greatly reduced. The application has the characteristics of simplicity, high efficiency and strong adaptability, and can make the maritime unmanned target identification system more accurately capture and identify targets.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing technology, and relates to a target recognition method for unmanned maritime platforms based on knowledge distillation and semi-supervised learning. Background Technology

[0002] Traditional methods for target identification in marine environments typically rely on deep learning models, which often have a large number of parameters and require extensive labeled data for training. However, due to the complexity and variability of the marine environment, obtaining high-quality labeled data is extremely difficult and expensive. The existence of a large amount of unlabeled data makes labeling this data a very time-consuming and labor-intensive task.

[0003] To address the problem of insufficient labeled data, semi-supervised learning methods have gained increasing attention. Semi-supervised learning trains the model by combining a small amount of labeled data with a large amount of unlabeled data. The unlabeled data can be used to generate pseudo-labels through the model, thus effectively utilizing it during training and improving the model's recognition ability. The introduction of semi-supervised learning not only alleviates the pressure of data labeling but also makes full use of the existing large amount of unlabeled data, improving the overall performance of the model. Common applications of semi-supervised learning are mainly in classification problems, with limited application in object recognition tasks.

[0004] Besides the challenges of data labeling, model lightweighting is also an important research direction. Deploying efficient object recognition models on resource-constrained edge devices is crucial. Knowledge distillation is a technique for model lightweighting that improves the performance of a student model by transferring the knowledge of a pre-trained teacher model to the student model. Hinton et al. first proposed this method and demonstrated that guiding the student network through distillation loss (Kullback-Leibler divergence) can effectively improve its performance.

[0005] Knowledge distillation technology enables models to run efficiently on edge devices by reducing the number of model parameters while maintaining or improving model performance.

[0006] Image enhancement is a crucial technique for target recognition tasks, manipulating image data to improve the model's adaptability and recognition accuracy in various situations. Traditional image enhancement methods include geometric transformations (such as rotation, scaling, and translation), color space transformations (such as brightness adjustment, contrast enhancement, and color balancing), and pixel-level operations (such as sharpening, blurring, and noise addition). These techniques can significantly improve model performance in many scenarios; however, for maritime target recognition tasks, the effectiveness of traditional methods is not significant.

[0007] The shortcomings of these existing technologies all affect the accuracy of target identification by unmanned maritime platforms in complex environments, and make them difficult to deploy on edge devices. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a target recognition method for unmanned maritime platforms based on knowledge distillation and semi-supervised learning. By using rank-one prior image enhancement and semi-supervised learning to predict the location and category of objects in unlabeled images, this method solves the problems of low accuracy and limited labeled data in maritime target recognition. Furthermore, the knowledge distillation method addresses the issue of lightweight model deployment, ensuring the accuracy and real-time performance of maritime target detection and meeting practical application requirements.

[0009] To achieve the above objectives, this invention provides a target recognition method for unmanned maritime platforms based on knowledge distillation and semi-supervised learning, which includes the following steps:

[0010] Step 1: Analyze the problem of maritime target recognition, establish image data augmentation techniques, and obtain the original image after color space transformation (such as brightness adjustment and contrast enhancement) and rank-one prior image enhancement (ROP image enhancement).

[0011] Step 2: Through the effect analysis of image enhancement, a new image enhancement technique for target recognition - random occlusion - was established.

[0012] Redundancy is eliminated by randomly occluding multiple black blocks on the image. The combination of various data augmentation techniques improves the model's recognition capabilities under complex lighting and occlusion conditions.

[0013] Step 3: For target recognition, detailed information about targets of different sizes is captured through multi-mode feature fusion. A fully connected layer is used to map the feature map to the detection output. Each grid cell predicts multiple bounding boxes, the confidence score of each box, and the position of each box. Finally, post-processing (non-maximum suppression) is used to select the final bounding boxes, completing the target detection.

[0014] Step 4: Combining a small amount of labeled image data with a large amount of unlabeled image data, a semi-supervised learning-based object recognition network framework is adopted. For labeled image data, the bounding boxes, confidence scores, and object categories output by the network are compared with the true labels to calculate classification error, localization error, and confidence error, thereby updating the model.

[0015] Step 5: For unlabeled image data, strong and weak data augmentation are performed first. The image after weak augmentation is input into the target recognition network to obtain the output. The confidence of the pseudo-label is determined by a given threshold, thus completing the model training.

[0016] Step 6: For the trained target recognition model (teacher model), in order to make it lightweight for deployment on the edge, a knowledge distillation lightweighting technique is adopted. During the training process, the student model gradually learns the knowledge of the teacher model by minimizing the difference between its prediction results and the soft labels of the teacher model. This knowledge transfer mechanism enables the student model to improve its feature representation and recognition capabilities. Finally, a lightweight target recognition model is obtained.

[0017] The present invention has the following beneficial effects:

[0018] First, to further enhance the model's recognition and semi-supervised learning capabilities, color space transformation (such as brightness adjustment and contrast enhancement), rank-one prior image enhancement (ROP image enhancement), and random occlusion enhancement techniques were established. The combination of these techniques improved the model's recognition ability under complex lighting and occlusion conditions. Second, through semi-supervised learning, training was effectively combined with a small amount of labeled image data and a large amount of unlabeled image data. This enabled the model to achieve recognition capabilities approaching those of supervised learning, even with limited labeled image data. Finally, by introducing knowledge distillation technology, the knowledge taught to the model was effectively transferred, significantly improving the recognition ability of the student model while drastically reducing its parameter count. This invention is simple, efficient, and highly adaptable, enabling maritime unmanned target recognition systems to more accurately capture and identify targets. Attached Figure Description

[0019] Figure 1 This is a flowchart of the target recognition method for unmanned maritime platforms based on knowledge distillation and semi-supervised learning, as described in this invention.

[0020] Figure 2 This is a schematic diagram illustrating the principle of the knowledge distillation process in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0022] like Figure 1 , Figure 2 As shown, a target recognition method for unmanned maritime platforms based on knowledge distillation and semi-supervised learning includes the following steps:

[0023] Step 1: This invention employs color space transformation (such as brightness adjustment and contrast enhancement) and rank-one prior image enhancement (ROP image enhancement). ROP is a novel real-time scene restoration technique used to restore degraded images under different weather / imaging conditions, such as late night and hazy weather;

[0024] Step 2: Establish a new image augmentation technique for target recognition – random occlusion. Random occlusion eliminates information redundancy by randomly occluding multiple black blocks on an image. The combination of multiple data augmentation techniques improves the model's recognition ability under complex lighting and occlusion conditions;

[0025] Step 3: The YOLO (You Only Look Once) method for object recognition transforms the object detection task into a single regression problem, enabling the simultaneous prediction of multiple object types and bounding box locations during a single propagation. Specifically, the YOLO model architecture employs a convolutional neural network (CNN) for feature extraction and captures detailed information about objects of different sizes through multi-mode feature fusion. Fully connected layers map the feature maps to the detection output, with each grid cell predicting multiple bounding boxes, their confidence scores, and their locations. Finally, post-processing (non-maximum suppression) selects the final bounding boxes, completing object detection. The model's loss function combines classification error, localization error, and confidence error, defined as follows: Classification error measures the difference between the predicted class probability and the actual class. It uses a multi-class cross-entropy function, defined as:

[0026]

[0027] in, It is the number of grid cells. It is the first Categories in each grid The true probability, It is the first Categories in each grid The predicted probability, It is an indicator function, when the first... The value is 1 if a grid contains the target, and 0 otherwise.

[0028] Step 4: The positioning error is used to measure the difference between the predicted bounding box position and the actual position, and is defined as:

[0029] in, The first The center coordinates and the true values ​​of width and height of the bounding box in each grid. The first The predicted center coordinates and width and height of the bounding box in each grid. It is a hyperparameter that controls the weighting of positioning error;

[0030] Step 5: Confidence error measures the difference between the predicted bounding box confidence and the actual confidence, and is defined as:

[0031]

[0032] in, It is the first True confidence of the bounding box in each grid cell It is the first The prediction confidence of the bounding box in each grid cell. It is an indicator function, when the first... The value is 1 if a grid does not contain a target, and 0 otherwise. This is a hyperparameter that controls the weights of the confidence error for the targetless grid. In summary, the loss function of the YOLO model is: ;

[0033] Step 6: In the knowledge distillation process, the output of the teacher model is not directly applied, but rather passed to the student model as "soft labels." "Soft labels" refer to the teacher model's predictions of the input data; these predictions include not only the probability of the correct class but also the probability distribution of other classes. Compared to traditional hard labels, soft labels provide richer information. During training, the student model gradually learns the teacher model's knowledge by minimizing the difference between its predictions and the teacher model's soft labels. This process is achieved by minimizing the KL divergence loss function, as shown in the following formula:

[0034]

[0035] This loss function ensures that the student model's output is as close as possible to the teacher model's. This knowledge transfer mechanism enhances the student model's feature representation and recognition capabilities.

[0036] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the protection scope of the claims of the present invention.

Claims

1. A target recognition method for unmanned maritime platforms based on knowledge distillation and semi-supervised learning, characterized in that, It includes the following steps: Step 1: Employ color space transformation (such as brightness adjustment and contrast enhancement) and rank-one prior image enhancement (ROP image enhancement); ROP is a novel real-time scene restoration technique for restoring degraded images under different weather / imaging conditions; Step 2: Establish a new image enhancement technique for target recognition - random occlusion; random occlusion eliminates information redundancy by randomly occluding multiple black blocks on the image; the combination of multiple data augmentation techniques improves the model's recognition ability under complex lighting and occlusion conditions; Step 3: Implement the object recognition method using YOLO (You Only Look Once). This transforms the object detection task into a single-regression problem, simultaneously predicting the types and bounding box positions of multiple objects during a single propagation. Specifically, the YOLO model architecture uses a convolutional neural network (CNN) for feature extraction and captures detailed information about objects of different sizes through multi-mode feature fusion. Fully connected layers map the feature maps to the detection output, with each grid cell predicting multiple bounding boxes, their confidence scores, and their positions. Finally, post-processing, specifically non-maximum suppression, selects the final bounding boxes, completing the object detection. The model's loss function combines classification error, localization error, and confidence error, defined as follows: Classification error measures the difference between the predicted class probability and the actual class; a multi-class cross-entropy function is used, defined as: in, It is the number of grid cells. It is the first Categories in each grid The true probability, It is the first Categories in each grid The predicted probability, It is an indicator function, when the first... The value is 1 if a grid contains the target, and 0 otherwise. Step 4: The positioning error is used to measure the difference between the predicted bounding box position and the actual position, and is defined as: in, The first The center coordinates and the true values ​​of width and height of the bounding box in each grid. The first The predicted center coordinates and width and height of the bounding box in each grid. It is a hyperparameter that controls the weighting of positioning error; Step 5: Confidence error is used to measure the difference between the predicted bounding box confidence and the actual confidence, and is defined as: in, It is the first True confidence of the bounding box in each grid cell It is the first The prediction confidence of the bounding box in each grid cell. It is an indicator function, when the first... The value is 1 if a grid does not contain a target, and 0 otherwise. These are hyperparameters that control the weights of the confidence error in the targetless grid; the loss function of the YOLO model is: ; Step 6: In the knowledge distillation process, the output of the teacher model is not directly applied, but is passed to the student model as soft labels. Soft labels refer to the teacher model's predictions of the input data, which include not only the probability of the correct class but also the probability distribution of other classes. Compared with traditional hard labels, soft labels provide richer information. During training, the student model gradually learns the knowledge of the teacher model by minimizing the difference between its predictions and the teacher model's soft labels. This process is achieved by minimizing the KL divergence loss function, as shown in the following formula: This loss function makes the output of the student model as close as possible to the output of the teacher model; this knowledge transfer mechanism enables the student model to improve its feature representation and recognition capabilities.

Images