Mixed-flow water turbine runner crack and cavitation defect identification method

Abstract

The invention provides a mixed-flow water turbine runner crack and cavitation defect identification method which comprises the following steps: acquiring a normal image, a crack image and a cavitation image of a water turbine runner, and processing by using an image enhancement technology to generate a water turbine runner image data set; constructing a water turbine runner crack and cavitation defect identification network model based on the improved YOLOv8; training a water turbine runner crack and cavitation defect identification network model based on the improved YOLOv8; and detecting the turbine runner defects by using the trained turbine runner crack and cavitation defect identification network model. According to the invention, computer vision and deep learning technologies are adopted, and the crack and cavitation defects of the turbine runner blade can be accurately detected.

Description

Technical Field

[0001] The invention relates to the technical field of water turbine runner monitoring, and in particular to a method for identifying cracks and cavitation defects of a Francis turbine runner. Background Art

[0002] As an important part of the hydroelectric generator set, the performance and reliability of the turbine runner directly affect the operation of the entire power generation system. During the operation of the hydroelectric unit, the turbine runner blades are often affected by a variety of unstable factors, including the transient force of the water flow, the erosion of solid particles in the water flow, and the operation of the runner under unstable conditions. These complex working conditions and external environment can easily cause defects such as cracks and cavitation in the runner blades. These defects will not only reduce the power generation efficiency of the unit, but also pose a serious threat to the safety and long-term reliability of the unit. The traditional method of turbine runner defect detection mainly relies on manual visual inspection, but this method has significant deficiencies in accuracy, efficiency and safety. Therefore, how to effectively and accurately detect cracks and cavitation defects in turbine runner blades has become an important issue that needs to be solved urgently.

[0003] With the rapid development of science and technology, especially computer vision and deep learning technology, methods based on image processing and deep learning are gradually being applied to the field of industrial defect detection. These methods can not only extract subtle defect features, but also effectively handle complex backgrounds and improve the accuracy and efficiency of detection. Summary of the invention

[0004] The present invention aims to provide a method for identifying cracks and cavitation defects in a Francis turbine runner, so as to accurately detect surface defects of the turbine runner by using computer vision and deep learning technology.

[0005] The present invention provides a method for identifying cracks and cavitation defects in a Francis turbine runner, comprising the following steps:

[0006] Collect normal images, crack images and cavitation images of the turbine runner, and process them using image enhancement technology to generate a turbine runner image dataset;

[0007] Construct a network model for turbine runner crack and cavitation defect identification based on improved YOLOv8;

[0008] Train a network model for turbine runner crack and cavitation defect recognition based on improved YOLOv8;

[0009] The trained turbine runner crack and cavitation defect recognition network model is used to detect turbine runner defects.

[0010] In some embodiments, the processing using image enhancement technology includes:

[0011] The collected normal images, crack images and cavitation images of the turbine runner are processed by using image enhancement techniques such as image cropping, image flipping and image blurring to expand the image data;

[0012] Annotation tools are used to annotate each image so that each image has a corresponding label file, and finally a turbine runner image dataset is generated after image processing.

[0013] In some embodiments, the turbine runner image dataset is divided into a training set, a validation set, and a test set according to a certain ratio.

[0014] In some embodiments, the turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 includes a backbone network Backbone, a neck network Neck and a head network Head.

[0015] In some embodiments, the backbone network Backbone includes 5 Conv modules, 4 C2f modules and 1 SPPCSPC+SimAM module, and the connection relationship is: the first Conv module is connected to the SPPCSPC+SimAM module in sequence through 4 feature extraction stages;

[0016] The four feature extraction stages are respectively the first feature extraction stage, the second feature extraction stage, the third feature extraction stage, and the fourth feature extraction stage; the first feature extraction stage includes the second Conv module and the first C2f module, the second feature extraction stage includes the third Conv module and the second C2f module, the third feature extraction stage includes the fourth Conv module and the third C2f module, and the fourth feature extraction stage includes the fifth Conv module and the fourth C2f module.

[0017] In some embodiments, the structure of the SPPCSPC+SimAM module is: the input end is connected to the sixth Conv module, the first SimAM module, the MaxPool layer, the first Concat module, the seventh Conv module, the eighth Conv module, the second SimAM module, the second Concat module and the ninth Conv module in sequence on the one hand, and is connected to the second SimAM module via the tenth Conv module on the other hand.

[0018] In some embodiments, the neck network Neck includes 3 Conv modules, 6 C2f modules, 3 UnSample modules, 6 ConCat modules and 4 GAM modules, and the connection relationship is as follows: the first UnSample module is connected to the third ConCat module, the fifth C2f module, the second UnSample module, the fourth ConCat module, the sixth C2f module, the first GAM module, the third UnSample module, the fifth ConCat module, the seventh C2f module, the second GAM module, the eleventh Conv module, the sixth ConCat module, the eighth C2f module, the third GAM module, the twelfth Conv module, the seventh ConCat module, the ninth C2f module, the fourth GAM module, the thirteenth Conv module, the eighth ConCat module, and the tenth C2f module; wherein the fifth C2f module is connected to the seventh ConCat module, the first GAM module is connected to the sixth ConCat module, the third ConCat module is connected to the third C2f module of the backbone network Backbone, the fourth ConCat module is connected to the second C2f module of the backbone network Backbone, and the fifth ConCat module is connected to the first C2f module of the backbone network Backbone.

[0019] In some embodiments, the head network Head includes four Detect modules, namely a first Detect module, a second Detect module, a third Detect module and a fourth Detect module, and their connection relationship is: the first Detect module is connected to the second GAM module of the neck network Neck, the second Detect module is connected to the third GAM module of the neck network Neck, the third Detect module is connected to the fourth GAM module of the neck network Neck, and the fourth Detect module is connected to the tenth C2f module of the neck network Neck.

[0020] In some embodiments, the MPDIoU loss function is used to train a turbine runner crack and cavitation defect recognition network model based on improved YOLOv8.

[0021] In some embodiments, the MPDIoU loss function is expressed as:

[0022] Loss = 1-MPDIoU (1)

[0023]

[0024]

[0025] In formula (1) to formula (6), and Represent the center point coordinates of the real annotation bounding box and the predicted bounding box, respectively, gt and h gt represents the width and height of the ground-truth bounding box, w prd and h prd Represents the width and height of the prediction box.

[0026] In summary, the present invention uses computer vision and deep learning technology to accurately detect cracks and cavitation defects in turbine runner blades. It also has the following specific improvements and advantages:

[0027] 1. Since the proportion of images containing defective areas in the collected turbine runner images is relatively small, if the original data is directly used for training, the model may be underfitted. Therefore, the collected defective images are first expanded using image enhancement techniques such as image cropping, image flipping, and image blurring, and the processed images are used as training data sets. A turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 is constructed, and a turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 is trained and generated for detecting turbine runner defects. The method proposed in the present invention has high accuracy for turbine runner defect detection.

[0028] 2. A turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 is adopted. The improvements are: (1) adding a SPPCSPC+SimAM module to the backbone network of the YOLOv8 network to reduce feature loss; (2) introducing a GAM module in the neck network of the YOLOv8 network to improve the expression ability of the model; (3) adding a high-resolution detection head to the head network of the YOLOv8 network to enhance the model's ability to detect tiny targets; (4) replacing the original CIoU of the YOLOv8 network with MPDIoU as a new loss function to improve the convergence speed and regression accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The figure is an overall flow chart of the method for identifying cracks and cavitation defects in a Francis turbine runner according to an embodiment of the present invention.

[0030] Figure 2 It is a structural schematic diagram of a turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 in an embodiment of the present invention.

[0031] Figure 3 It is a structural schematic diagram of the SimAM module in an embodiment of the present invention.

[0032] Figure 4 Schematic diagram of the structure of the C2f module in an embodiment of the present invention.

[0033] Figure 5 It is a structural schematic diagram of the SPPCSPC+SimAM module in an embodiment of the present invention.

[0034] Figure 6 Schematic diagram of the structure of the GAM module in an embodiment of the present invention.

[0035] Figure 7 A schematic diagram of a turbine runner crack defect detection result provided in an embodiment of the present invention.

[0036] Figure 8 A schematic diagram of the detection results of cavitation defects in a turbine runner provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0038] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Example

[0040] like Figure 1 As shown, this embodiment provides a method for identifying cracks and cavitation defects in a Francis turbine runner, comprising the following steps:

[0041] S1, collecting normal images, crack images and cavitation images of a turbine runner, and processing them using image enhancement technology to generate a turbine runner image data set;

[0042] In some embodiments, the processing using image enhancement technology includes: first processing the collected normal images, crack images and cavitation images of the turbine runner using image enhancement technologies such as image cropping, image flipping, and image blurring to expand the image data, and at the same time using a labeling tool (such as open source Labelme) to label each image so that each image has a corresponding label file, and finally generating a turbine runner image data set after image processing.

[0043] In some embodiments, in order to divide the turbine runner image data set into a training set, a validation set, and a test set according to a certain ratio, it is convenient to subsequently train, validate, and test the turbine runner crack and cavitation defect recognition network model based on the improved YOLOv8. The division ratio can be set according to actual needs, for example, the ratio of the training set, the validation set, and the test set is 6:2:2.

[0044] S2. Construct a network model for identifying cracks and cavitation defects in turbine runners based on improved YOLOv8;

[0045] In some embodiments, Figure 2 As shown, the turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 includes a backbone network Backbone, a neck network Neck and a head network Head.

[0046] S21. Build the backbone network:

[0047] The backbone network Backbone is mainly used to extract high-level features from the original input image. These features contain the global context and abstract information of the image. It converts the input image into a feature map and gradually extracts more abstract features, so that the network can better understand the image content.

[0048] In some embodiments, Figure 2 As shown, the backbone network Backbone includes 5 Conv modules, 4 C2f modules and 1 SPPCSPC+SimAM module, and the connection relationship is: the first Conv module is connected to the SPPCSPC+SimAM module in sequence through 4 feature extraction stages; each of the feature extraction stages is also a Conv-C2f combination stage, that is, each of the feature extraction stages includes a Conv module and a C2f module, thus, the 4 feature extraction stages are respectively the first feature extraction stage, the second feature extraction stage, the third feature extraction stage, and the fourth feature extraction stage; the first feature extraction stage includes the second Conv module and the first C2f module, the second feature extraction stage includes the third Conv module and the second C2f module, the third feature extraction stage includes the fourth Conv module and the third C2f module, and the fourth feature extraction stage includes the fifth Conv module and the fourth C2f module. Wherein:

[0049] The Conv module is a combination of convolution-batch normalization-activation function, which is the most basic operation unit in the turbine runner crack and cavitation defect recognition model based on improved YOLOv8, and is used to extract and process features;

[0050] The C2f module is a cross-scaling path structure and a lightweight module used to enhance the gradient flow and multi-scale information of features. Its structure is as follows Figure 4 shown.

[0051] The principle of the backbone network Backbone is as follows:

[0052] S211, the input image is converted into a primary feature map through the first Conv module;

[0053] S222, the primary feature map is sequentially subjected to four feature extraction stages to extract deep image features, wherein the four feature extraction stages correspond to four detection heads respectively, capturing feature information at different levels;

[0054] S223. The depth image features are passed through the SPPCSPC+SimAM module to reduce feature loss and facilitate the fusion of different scale features of the neck network Neck.

[0055] The SPPCSPC+SimAM module is a structure composed of the SPPCSPC module and the SimAM module. It is used to perform multi-scale pooling on feature maps to enhance the receptive field and robustness of the features. Figure 3 As shown in the figure, the SimAM module is a simple, efficient, and lightweight three-dimensional attention module that infers three-dimensional (taking into account both spatial and channel dimension correlation) attention weights through feature maps in the feature layer without adding parameters to the original network. Adding the SimAM module to the network will not increase the complexity of the network, but can increase the representation capability of the original network to a certain extent, and will not affect the training and reasoning speed of the network.

[0056] like Figure 5 As shown, in the embodiment of the present invention, the SimAM module is used to modify the SPPCSPC module: first, the two Conv modules located before the MaxPool layer in the original SPPCSPC module are cropped to reduce the filtering of the Conv module on the edge information of the small target, and at the same time, the calculation amount of the network model is also reduced. Then, a SimAM module is introduced at the cropped position to solve the problem of insufficient extraction of small target defect features. While dividing the pixels of the complex background area and the small target area, the SimAM module assigns 3D attention weights to each neuron to enhance the sensitivity of the complex background and small target areas. At the same time, adding another SimAM module before the MaxPool layer and before the last Concat module helps to divide the defect area from the non-defect area, enhance the network model's attention to the small target area, reduce the loss of features, and improve the robustness of the network model by reducing the impact of the complex background on the defects.

[0057] Therefore, the structure of the SPPCSPC+SimAM module in this embodiment is as follows: the input end is connected to the sixth Conv module, the first SimAM module, the MaxPool layer, the first Concat module, the seventh Conv module, the eighth Conv module, the second SimAM module, the second Concat module and the ninth Conv module in sequence on the one hand, and is connected to the second SimAM module via the tenth Conv module on the other hand. Among them, the ConCat module is a feature splicing operation, which is used to fuse features at different levels to adapt to target detection of different scales.

[0058] S22. Build the Neck network:

[0059] The neck network Neck is mainly used to further process features based on the features extracted by the backbone network, realize feature fusion, refinement and extraction, and help to effectively fuse multi-scale and multi-channel features to improve the detection performance of targets of different scales and sizes.

[0060] In some embodiments, Figure 2 As shown, the neck network Neck includes 3 Conv modules, 6 C2f modules, 3 UnSample modules, 6 ConCat modules and 4 GAM modules, and the connection relationship is as follows: the first UnSample module is sequentially connected to the third ConCat module, the fifth C2f module, the second UnSample module, the fourth ConCat module, the sixth C2f module, the first GAM module, the third UnSample module, the fifth ConCat module, the seventh C2f module, the second GAM module, the eleventh Conv module, the sixth ConCat module, and the eighth C2f module , the third GAM module, the twelfth Conv module, the seventh ConCat module, the ninth C2f module, the fourth GAM module, the thirteenth Conv module, the eighth ConCat module, and the tenth C2f module; wherein the fifth C2f module is connected to the seventh ConCat module, the first GAM module is connected to the sixth ConCat module, the third ConCat module is connected to the third C2f module of the backbone network Backbone, the fourth ConCat module is connected to the second C2f module of the backbone network Backbone, and the fifth ConCat module is connected to the first C2f module of the backbone network Backbone.

[0061] The neck network Neck adopts a typical feature pyramid network structure, namely the FPN+PAN structure. The FPN layer conveys strong semantic features from top to bottom, and the PAN layer conveys strong positioning features from bottom to top. The two work together to fuse parameters of different detection layers from different backbone layers, so that the location information and category information of the target are retained to the greatest extent. In the neck network Neck, the UnSample module is an upsampling operation, which is used to enlarge the low-resolution feature map to high resolution for more precise target detection. At the same time, the neck network Neck adds 4 GAM modules based on the original YOLOv8 network model. The GAM module is an attention mechanism module, which can adaptively adjust the feature map in the spatial dimension and channel dimension to improve the performance and generalization ability of the network model. Figure 6 As shown in the figure, the GAM module contains a channel attention submodule and a spatial attention submodule. The channel attention submodule usually uses three-dimensional arrangement and multi-layer perceptron to process the channel information of the feature map, and enhances or suppresses the features of a specific channel by learning the dependencies between different channels; the spatial attention submodule uses structures such as convolutional layers to process the spatial information of the feature map, generates a spatial attention map by learning the importance of each spatial position, and fuses it with the original feature map to refine the representation of the feature. The GAM module realizes cross-dimensional interaction by processing information in both channel and spatial dimensions at the same time, so that the key information in the input data can be captured more comprehensively. Introducing the GAM module in the neck network Neck of the YOLOv8 model can enhance the model's fusion of target features to improve the model's expression ability and thus improve the accuracy of detection.

[0062] S23. Build the head network Head:

[0063] The head network Head is mainly used to detect and identify targets on the fused feature map. A high-resolution detection head is added on the basis of the original YOLOv8 network model to enhance the model's ability to detect tiny targets, thereby obtaining smaller feature information of the defective area of ​​the turbine runner.

[0064] In some embodiments, Figure 2 As shown, the head network Head includes 4 Detect modules, namely the first Detect module, the second Detect module, the third Detect module and the fourth Detect module, and their connection relationship is: the first Detect module is connected to the second GAM module of the neck network Neck, the second Detect module is connected to the third GAM module of the neck network Neck, the third Detect module is connected to the fourth GAM module of the neck network Neck, and the fourth Detect module is connected to the tenth C2f module of the neck network Neck.

[0065] S3, training a network model for turbine runner crack and cavitation defect recognition based on improved YOLOv8;

[0066] The turbine runner image data set is input into the turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 for training, and the network weights obtained by training are saved to obtain a trained turbine runner crack and cavitation defect recognition network model for identifying turbine runner crack defects and cavitation defects.

[0067] The YOLOv8 network model uses CIoU (Complete Intersection over Union) as the loss function, which integrates three important geometric factors: overlapping area, centroid distance, and aspect ratio, to improve the positioning ability of the predicted box. However, when the predicted box has the same aspect ratio as the true box, but the width and height values ​​are completely different, it may hinder the model from effectively optimizing the similarity between the two boxes, so CIoU still has a lot of room for improvement. In order to solve this problem, MPDIoU is selected in the embodiment of the present invention to replace CIoU in the original network. MPDIoU simplifies the comparison of the similarity between the true box and the predicted box, and can improve the convergence speed and regression accuracy. Its calculation formula is shown in formulas (1)-(6):

[0068] Loss = 1-MPDIoU (1)

[0069]

[0070] In formula (1) to formula (6), and Represent the center point coordinates of the real annotation bounding box and the predicted bounding box, respectively, gt and h gt represents the width and height of the ground-truth bounding box, w prd and h prd Represents the width and height of the prediction box.

[0071] S4. Use the trained turbine runner crack and cavitation defect recognition network model to detect turbine runner defects.

[0072] like Figure 7 As shown, the crack image detection result of the turbine runner blade surface is marked with a small red frame on the crack part, and the pinyin abbreviation "LW" is used to indicate that the image is a crack image, and the decimal "0.98" is used to indicate that the probability that the image is a crack image is 98%.

[0073] like Figure 8As shown in the figure, the cavitation image detection result of the turbine runner blade surface is marked with a small blue box in the crack part, and the pinyin abbreviation "QS" is used to indicate that the image is a cavitation image, and the decimal "0.96" is used to indicate that the probability that the image is a crack image is 96%. When there are multiple cavitations in the image, all cavitation areas will be marked.

[0074] It should be noted that the above-mentioned marking method is only an example in this embodiment, and the marked box, color, value, etc. can be designed according to actual needs.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying cracks and cavitation defects in a Francis turbine runner, characterized in that: The steps include: Collect normal images, crack images and cavitation images of the turbine runner, and process them using image enhancement technology to generate a turbine runner image dataset; Construct a network model for turbine runner crack and cavitation defect identification based on improved YOLOv8; Train a network model for turbine runner crack and cavitation defect recognition based on improved YOLOv8; The trained turbine runner crack and cavitation defect recognition network model is used to detect turbine runner defects.

2. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 1, characterized in that: The processing using the image enhancement technology includes: The collected normal images, crack images and cavitation images of the turbine runner are processed by using image enhancement techniques such as image cropping, image flipping and image blurring to expand the image data; Annotation tools are used to annotate each image so that each image has a corresponding label file, and finally a turbine runner image dataset is generated after image processing.

3. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 1, characterized in that: The turbine runner image data set is divided into a training set, a validation set and a test set according to a certain ratio.

4. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 1, characterized in that: The turbine runner crack and cavitation defect recognition network model based on improved YOLOv8 includes a backbone network Backbone, a neck network Neck and a head network Head.

5. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 4, characterized in that: The backbone network Backbone includes 5 Conv modules, 4 C2f modules and 1 SPPCSPC+SimAM module, and the connection relationship is as follows: the first Conv module is connected to the SPPCSPC+SimAM module in sequence through 4 feature extraction stages; The four feature extraction stages are respectively the first feature extraction stage, the second feature extraction stage, the third feature extraction stage, and the fourth feature extraction stage; the first feature extraction stage includes the second Conv module and the first C2f module, the second feature extraction stage includes the third Conv module and the second C2f module, the third feature extraction stage includes the fourth Conv module and the third C2f module, and the fourth feature extraction stage includes the fifth Conv module and the fourth C2f module.

6. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 5, characterized in that: The structure of the SPPCSPC+SimAM module is: the input end is connected to the sixth Conv module, the first SimAM module, the MaxPool layer, the first Concat module, the seventh Conv module, the eighth Conv module, the second SimAM module, the second Concat module and the ninth Conv module in sequence on the one hand, and is connected to the second SimAM module via the tenth Conv module on the other hand.

7. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 5, characterized in that: The neck network Neck includes 3 Conv modules, 6 C2f modules, 3 UnSample modules, 6 ConCat modules and 4 GAM modules, and the connection relationship is as follows: the first UnSample module is connected to the third ConCat module, the fifth C2f module, the second UnSample module, the fourth ConCat module, the sixth C2f module, the first GAM module, the third UnSample module, the fifth ConCat module, the seventh C2f module, the second GAM module, the eleventh Conv module, the sixth ConCat module, the eighth C2f module, the first GAM module, the third UnSample module, the fifth ConCat module, the seventh C2f module, the second GAM module, the eleventh Conv module, the sixth ConCat module, the eighth C2f module, the first GAM module, the third The third GAM module, the twelfth Conv module, the seventh ConCat module, the ninth C2f module, the fourth GAM module, the thirteenth Conv module, the eighth ConCat module, and the tenth C2f module; among them, the fifth C2f module is connected to the seventh ConCat module, the first GAM module is connected to the sixth ConCat module, the third ConCat module is connected to the third C2f module of the backbone network Backbone, the fourth ConCat module is connected to the second C2f module of the backbone network Backbone, and the fifth ConCat module is connected to the first C2f module of the backbone network Backbone.

8. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 7, characterized in that: The head network Head includes four Detect modules, namely the first Detect module, the second Detect module, the third Detect module and the fourth Detect module, and their connection relationship is: the first Detect module is connected to the second GAM module of the neck network Neck, the second Detect module is connected to the third GAM module of the neck network Neck, the third Detect module is connected to the fourth GAM module of the neck network Neck, and the fourth Detect module is connected to the tenth C2f module of the neck network Neck.

9. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 1, characterized in that: The MPDIoU loss function is used to train the turbine runner crack and cavitation defect recognition network model based on improved YOLOv8.

10. The method for identifying cracks and cavitation defects in a Francis turbine runner according to claim 9, characterized in that: The MPDIoU loss function is expressed as: Loss = 1-MPDIoU (1) In formula (1) to formula (6), and ) represent the center point coordinates of the real annotation bounding box and the predicted bounding box, respectively, and w gt and h gt represents the width and height of the ground-truth bounding box, w prd and h prd Represents the width and height of the prediction box.

Images