Deep learning-based inland ship target detection and identification method

Abstract

The invention provides an inland ship target detection and identification method based on deep learning, and the method comprises the steps: building a multi-layer shallow information network through an improved YOLOv8 algorithm, adjusting a detection layer, adding a shallow detection head, capturing the features of a small target and the semantic information of the small target more finely, enhancing the target identification capability in a targeted manner, and improving the detection efficiency. The function of conveniently and interactively detecting and identifying the inland ship target is realized. According to the method, an EfficientNetV2 lightweight network optimization strategy is combined, a model network structure is improved, and a more effective loss function is adopted, so that the calculation efficiency and precision are improved, and the model training and reasoning efficiency is guaranteed. A dynamic non-monotonic focusing mechanism is adopted to ensure that the anchor frame is more fit with a real bounding box model, and the overall performance of the detector is improved. According to the method, the problems that the detection effect of the small target is not ideal, the detection anchor frame and the actual boundary are not tightly attached, and the system interaction is insufficient in a complex water surface environment are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent navigation assistance, and specifically relates to a method for detecting and identifying inland river vessel targets based on deep learning. Background Art

[0002] At present, water traffic is becoming increasingly busy, and the needs for safety supervision, traffic flow management, and emergency response of inland waterway traffic are becoming increasingly prominent. As one of the key technologies to achieve these functions, ship detection and recognition technology has received extensive attention and research in recent years. This technology is one of the core tasks in computer vision, which involves identifying and locating objects of interest from images or videos. Traditional detection methods mostly rely on manual monitoring or signal analysis based on radar, AIS (Automatic Identification System) and other equipment. These methods have limitations such as high cost, poor real-time performance, and susceptibility to environmental interference.

[0003] Today, the rapid development of deep learning technology has brought revolutionary changes to the field of ship detection and identification. As a leader in the detection field, the YOLO (You Only Look Once) series of algorithms have broad applications in areas such as navigation safety and traffic management, and are the foundation for the future development of intelligent navigation and maritime supervision.

[0004] However, the detection capabilities of the traditional YOLOv8 still have some limitations and challenges. The traditional YOLOv8 algorithm performs poorly when detecting small targets, especially when the image resolution is low or the contrast between the object and the background is not high; for example, when detecting small objects in the image, the accuracy of the algorithm may be affected, it is difficult to accurately distinguish and identify in complex or changing backgrounds, the algorithm performance will decrease in the edge area of ​​the image, and the detection anchor box is not closely aligned with the actual position; the lack of an intuitive and easy-to-use user interface makes it difficult for non-professional users to set up and adjust parameters of the algorithm, and it is difficult for non-professional users to fully utilize the potential of the algorithm. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a deep learning-based inland river vessel target detection and recognition method for convenient and interactive detection and recognition of inland river vessel targets.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: a method for detecting and identifying inland river vessel targets based on deep learning, comprising the following steps:

[0007] S1: Build a multi-layer network structure model, including a shallow small target detection head for capturing shallow network features and extracting output image feature information;

[0008] S2: Combined with the EfficientNetV2 lightweight network optimization strategy, the stem module, MBConv module and Fused-MBConv module are introduced into the model to improve the network structure of the model;

[0009] S3: Based on the WIoU loss function, a dynamic non-monotonic focusing mechanism is adopted to enable the model to achieve dynamically changing gradient gain;

[0010] S4: The shipping image data collected on the inland river is used to form an inland river dataset for training the model.

[0011] According to the above scheme, in step S1, the specific steps are:

[0012] The multi-level network structure model includes multiple shallow networks;

[0013] Shallow networks are used to pass feature information to the model to add new detection layers;

[0014] The shallow network includes a 160*160 shallow detection layer P2, which is used to increase the overlap between the receptive field area and the detected area to finely capture features and semantic information and improve recognition capabilities.

[0015] According to the above scheme, in step S2, the specific steps of the Stems module processing data are:

[0016] First, the input image is pre-operated, including downsampling or padding, to fit the expected input size of the model;

[0017] Then the pre-operated image features are extracted and the spatial dimensions are transformed;

[0018] Normalize the image data after converting the spatial dimension to speed up the convergence and stabilize the training process;

[0019] Finally, the ReLU nonlinear activation function is used to enhance the expressiveness of the model.

[0020] According to the above scheme, in step S2, the specific steps of the MBConv module processing data are:

[0021] First, we apply a 1*1 convolutional layer to transform the features, expanding the number of channels to four times the original value.

[0022] Then 3*3 depth separation convolution is used to process features to reduce the number of parameters;

[0023] The SE module models the interdependence between feature channels and dynamically adjusts and calibrates the importance of each feature channel in the network, thereby making more efficient use of computing resources and improving accuracy.

[0024] Finally, the number of channels is restored to its initial size to ensure that the output layer of the module has the same dimension as the input layer, and the output layer of the module is feature fused with the input layer.

[0025] Furthermore, in step S2, the Fused-MBConv module uses 3*3 convolution to process large-size features to reduce the amount of calculation and the amount of parameters.

[0026] According to the above scheme, in step S3, the specific steps are:

[0027] Let x and y represent the coordinates of the center of the anchor box, x gt ,y gt represents the coordinates of the center of the ground-truth bounding box, W g , H g Respectively represent the length and width of the minimum enclosing box; set Increase the loss size of normal quality anchor boxes; set Reduce the high quality anchor box And when the anchor box and the true bounding box overlap well, the focus on the distance of the center point is significantly reduced:

[0028]

[0029] Furthermore, in step S3, it is assumed that is an exponential moving average with momentum m, is the monotonic focusing coefficient when γ = 1 The dynamic non-monotonic focusing mechanism uses "outlier degree β" to replace IoU to evaluate the quality of the anchor box:

[0030]

[0031] Implement a gradient gain distribution strategy:

[0032] Lower the priority of anchor boxes with small outliers to avoid over-optimization;

[0033] For anchor boxes with large outliers, the priority is lowered to reduce the negative impact.

[0034] Furthermore, in step S3, let r represent the gradient gain, and α and δ be hyperparameters; the gradient gain allocation strategy of the anchor frame is dynamically adjusted according to the actual situation during the training process to improve the robustness and accuracy of the model:

[0035]

[0036] According to the above scheme, the following steps are also included:

[0037] S5: Use the interface display framework PyQt5 to develop the system display interface.

[0038] A computer memory stores a computer program that can be executed by a computer processor, and the computer program executes an inland vessel target detection and recognition method based on deep learning.

[0039] The beneficial effects of the present invention are:

[0040] 1. The present invention discloses a method for detecting and identifying inland river vessel targets based on deep learning. By improving the YOLOv8 algorithm, a multi-level shallow information network is constructed, the detection layer is adjusted, and a shallow detection head is added to more finely capture the features of small targets and their semantic information, thereby enhancing the target recognition capability in a targeted manner and realizing the function of convenient and interactive detection and identification of inland river vessel targets.

[0041] 2. This invention combines the EfficientNetV2 lightweight network optimization strategy, improves the model network structure and adopts a more effective loss function, thereby improving the computing efficiency and accuracy and ensuring the efficiency of model training and reasoning.

[0042] 3. The present invention adopts a dynamic non-monotonic focusing mechanism to ensure that the anchor box is more closely aligned with the real bounding box model, thereby improving the overall performance of the detector.

[0043] 4. The present invention solves the problems of unsatisfactory detection of small targets, insufficient fit between the detection anchor frame and the actual boundary, and insufficient system interactivity in complex water surface environments.

[0044] 5. The present invention forms an intuitive display interface and interactive interface, providing an important reference for intelligent navigation research and safety supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of an embodiment of the present invention.

[0046] Figure 2 It is a structural diagram of a shallow-level detection head according to an embodiment of the present invention.

[0047] Figure 3 It is a schematic diagram of the overall network structure of an embodiment of the present invention.

[0048] Figure 4 Schematic diagram of the WIoU loss function of an embodiment of the present invention.

[0049] Figure 5 It is a schematic diagram of a system interface of an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0051] See also Figure 1, an embodiment of the present invention comprises the following steps:

[0052] S1: Build a multi-layer network structure model, add a shallow small target sub-detection head, capture the shallower network layer P2 layer features, and extract the image feature information output by the network. The shallower network layer retains more spatial resolution and contains more spatial information, which is used to pass feature information to the model to add a new detection layer.

[0053] The multi-level network structure model includes a multi-level shallow information network, combined with the EfficientNetV2 lightweight network optimization strategy to improve the network structure and WIoU loss function, self-built riverside data set and system interface display system. Figure 2 As shown in the figure, the multi-level shallow information network introduces a 160*160 shallow detection layer P2, which increases the overlap between the receptive field area and the detected area, thereby capturing its features and semantic information more finely and improving the recognition ability.

[0054] S2: Figure 3 As shown in the figure, combined with the EfficientNetV2 lightweight network optimization strategy, the three modules of stem, MBConv, and Fused-MBConv are introduced to improve the model network structure.

[0055] The Stems module is mainly used to perform preliminary image processing tasks and extract preliminary features from input image data. First, in order to adapt to the expected input size of the model, the image is downsampled or padded, and then features are extracted and spatial dimensions are converted; then normalization is performed to achieve faster convergence and a more stable training process; finally, the ReLU nonlinear activation function is used to enhance the expressiveness of the model;

[0056] The advantage of the MBConv module is that it has fewer module parameters and lower FLOPs, which is more effective in the later stages. This module first applies a dimensional 1*1 convolution layer for feature conversion, expanding the number of channels to four times the original value; then uses a 3*3 depthwise separable convolution to further process the features and reduce the number of parameters; the SE (Squeeze Excitation) module is used to model the interdependence between feature channels, dynamically adjust and calibrate the importance of each feature channel in the network, so as to more efficiently utilize computing resources and improve accuracy; finally, the number of channels is restored to the initial size to ensure that the output of the module has the same dimension as the input, and the output of the module is feature-fused with its input layer;

[0057] Compared with the MBConv module, the main change of the Fused-MBConv module is that in the process of training images, the early layers have a better training speed for large-size training images, so conventional 3*3 convolution is used to process large-size feature maps to reduce the amount of calculation and the number of parameters.

[0058] S3: Figure 4 As shown in the figure, the Wise-IoU (Wise Intersection over Union, WIoU) loss function based on the dynamic non-monotonic focusing mechanism adopts a dynamic non-monotonic focusing mechanism to achieve dynamically changing gradient gain, which improves the performance by effectively handling the low-quality case problem in the bounding box.

[0059] set up This increases the loss size of the normal quality anchor box and sets Reduce the high quality anchor box And when the anchor box and the true bounding box overlap well, the focus on the distance of the center point is significantly reduced.

[0060]

[0061] x, y represents the ordinate of the center of the anchor box, x gt ,y gt represents the ordinate of the center of the ground-truth bounding box, W g , H g Indicates the length and width of the minimum enclosing box.

[0062] Furthermore, the dynamic mechanism uses "outlier β" to replace IoU to evaluate the quality of anchor frames, and implements an accurate gradient gain allocation strategy. For anchor frames with smaller outliers (i.e., higher quality anchor frames), their priority is reduced to avoid over-optimization; for anchor frames with larger outliers (i.e., lower quality anchor frames), their priority is also reduced to reduce their negative impact. This strategy enables the model to focus on anchor frames of common quality, thereby improving the overall performance of the detector.

[0063]

[0064] represents an exponential moving average with momentum m, represents the monotonic focusing coefficient when γ = 1

[0065] Furthermore, the focusing mechanism is dynamic, that is, the gradient gain allocation strategy of the anchor frame will be dynamically adjusted according to the actual situation during the training process, and can more flexibly cope with changes in anchor frames of different qualities, thereby improving the robustness and accuracy of the model.

[0066]

[0067] Among them, r represents the gradient gain, and α and δ are hyperparameters.

[0068] S4: A representative Yangtze River dataset is formed by collecting shipping image data along the Yangtze River for training this model.

[0069] S5: Use the interface display framework PyQt5 to develop the system display interface to achieve intuitive display of information and convenient interaction.

[0070] like Figure 5 As shown, the system display interface formed by the present invention can select different models, detection objects, etc. according to actual needs, which greatly increases the interactivity and provides an intuitive display interface.

[0071] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0072] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A method for inland river vessel target detection and recognition based on deep learning, characterized in that: The following steps are involved: S1: Build a multi-layer network structure model, including a shallow small target detection head for capturing shallow network features and extracting output image feature information; S2: Combined with the EfficientNetV2 lightweight network optimization strategy, the stem module, MBConv module and Fused-MBConv module are introduced into the model to improve the network structure of the model; S3: Based on the WIoU loss function, a dynamic non-monotonic focusing mechanism is adopted to enable the model to achieve dynamically changing gradient gain; S4: The shipping image data collected on the inland river is used to form an inland river dataset for training the model.

2. According to claim 1, a method for detecting and identifying inland river vessel targets based on deep learning is characterized in that: In the step S1, the specific steps are: The multi-level network structure model includes multiple shallow networks; Shallow networks are used to pass feature information to the model to add new detection layers; The shallow network includes a 160*160 shallow detection layer P2, which is used to increase the overlap between the receptive field area and the detected area to finely capture features and semantic information and improve recognition capabilities.

3. According to claim 1, a method for detecting and identifying inland river vessel targets based on deep learning is characterized in that: In step S2, the specific steps of processing data by the Stems module are: First, the input image is pre-operated, including downsampling or padding, to fit the expected input size of the model; Then the pre-operated image features are extracted and the spatial dimensions are transformed; Normalize the image data after converting the spatial dimension to speed up the convergence and stabilize the training process; Finally, the ReLU nonlinear activation function is used to enhance the expressiveness of the model.

4. The method for detecting and identifying inland river vessel targets based on deep learning according to claim 1, characterized in that: In step S2, the specific steps of the MBConv module processing data are: First, we apply a 1*1 convolutional layer to transform the features, expanding the number of channels to four times the original value. Then 3*3 depth separation convolution is used to process features to reduce the number of parameters; The SE module models the interdependence between feature channels and dynamically adjusts and calibrates the importance of each feature channel in the network, thereby making more efficient use of computing resources and improving accuracy. Finally, the number of channels is restored to its initial size to ensure that the output layer of the module has the same dimension as the input layer, and the output layer of the module is feature fused with the input layer.

5. The method for detecting and identifying inland river vessel targets based on deep learning according to claim 4 is characterized in that: In step S2, the Fused-MBConv module uses 3*3 convolution to process large-size features to reduce the amount of calculation and the amount of parameters.

6. The method for detecting and identifying inland river vessel targets based on deep learning according to claim 1, characterized in that: In the step S3, the specific steps are: Let x and y represent the coordinates of the center of the anchor box, x gt ,y gt represents the coordinates of the center of the ground-truth bounding box, W g , H g Respectively represent the length and width of the minimum enclosing box; set Increase the loss size of normal quality anchor boxes; set Reduce the high quality anchor box And when the anchor box and the true bounding box overlap well, the focus on the distance of the center point is significantly reduced:

7. The method for detecting and identifying inland river vessel targets based on deep learning according to claim 6 is characterized in that: In the step S3, it is assumed that is an exponential moving average with momentum m, is the monotonic focusing coefficient when γ = 1 The dynamic non-monotonic focusing mechanism uses "outlier β" to replace IoU to evaluate the quality of the anchor box: Implement a gradient gain distribution strategy: Lower the priority of anchor boxes with small outliers to avoid over-optimization; For anchor boxes with large outliers, the priority is lowered to reduce the negative impact.

8. The method for detecting and identifying inland river vessel targets based on deep learning according to claim 7 is characterized in that: In step S3, let r represent the gradient gain, and α and δ are hyperparameters; Dynamically adjust the gradient gain allocation strategy of the anchor box according to the actual situation during training to improve the robustness and accuracy of the model:

9. The method for detecting and identifying inland river vessel targets based on deep learning according to claim 1, characterized in that: The following steps are also included: S5: Use the interface display framework PyQt5 to develop the system display interface.

10. A computer memory, characterized in that: A computer program executable by a computer processor is stored therein, and the computer program executes a method for detecting and identifying inland vessel targets based on deep learning as described in any one of claims 1 to 9.

Images