A method for breaking down the plastic film used in peanut cultivation and a matching breaking down component.

Abstract

This invention discloses a method for breaking peanut seedling membranes and a corresponding membrane-breaking assembly. The assembly is mounted on a membrane-breaking machine support with a walking and cruising function. It includes an image acquisition camera, a YOLOv3 deep learning neural network structure, and a membrane-breaking blade with an adjustable circumferential cutting radius. Based on the characteristic that peanut seedling height and leaf area are proportionally related and relatively fixed during growth, the method preprocesses peanut images acquired by the camera and adds the peanut seedling height as expected data to the training set. The trained model estimates the peanut seedling height by calculating the leaf area. During recognition, the acquired image is input into the model, and then the driving device determines the coordinates based on the camera parameter calibration information, driving the membrane-breaking blade to break the membrane of seedlings that meet the breaking requirements. This invention ensures the scientific cultivation of peanut seedlings, significantly reduces worker time, lowers labor costs, and improves production efficiency.

Description

Technical Field

[0001] This invention belongs to the field of agricultural production automation, and in particular to a method for breaking the film in peanut cultivation. Background Technology

[0002] Peanuts are a major crop in my country with abundant yields and large-scale cultivation. Peanut planting requires mulching with plastic film, which needs to be removed when the peanuts reach a certain height. However, due to the varying growth rates of peanuts, centralized film removal is not feasible to ensure production quality. Current methods primarily involve manual film removal, which is time-consuming and labor-intensive. With the development of agricultural automation, there is increasing focus on production efficiency and labor costs, leading to the application of more and more cutting-edge technologies in agricultural production.

[0003] Most existing film-breaking machine technologies on the market rely on mechanical structure design, with research focusing on improving the freedom of movement of the film-breaking machine's traveling carriage and the control methods that combine the film-breaking machine with manual operation. However, mechanical film-breaking machines still require manual operation or intervention and have not fully achieved so-called "production automation." In particular, the design of the film-breaking blades is mostly conical, which can easily damage the blades. This film-breaking method is too dependent on manual operation, and the complex mechanical structure is difficult to design and manufacture, resulting in high costs. Even if some labor is saved in certain aspects, the high mechanical cost makes it difficult to mass-produce the machines and popularize them in real production activities. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a structure and method for an automatic peanut height identification film breaking machine for peanut production sites. This solves the problem of low production efficiency caused by the inability of manual film breaking. At the same time, GPU acceleration technology is used to improve the identification accuracy and ensure the scientific nature of peanut cultivation.

[0005] To achieve the above objectives, the present invention provides a film-breaking machine for peanut cultivation, which is implemented using the following technical solution:

[0006] A method for breaking the film on a film-breaking machine component used in peanut cultivation, comprising the following steps:

[0007] (1) Input the peanut leaf image into the YOLOv3 neural network model for training. After the model detects, it outputs the leaf detection result. The detection result is whether the current growth status of the peanut plant meets the requirements for breaking the film.

[0008] (2) After the training is completed and the model is deployed, the peanut leaf recognition model will obtain the corresponding model file. At this time, the training results are converted into a format and then optimized by quantization before being deployed to the embedded MCU system. The embedded MCU system is connected to the film breaking machine component, which is mounted on the inter-row traveling vehicle.

[0009] (3) The film breaking machine component follows the traveling vehicle to collect images in real time. The neural network will output the judgment result based on the leaf image of the collected image. For plants that meet the requirements for film breaking, the system outputs the film breaking command, obtains the three-dimensional coordinate process, and executes the film breaking action. For leaves that do not meet the requirements for film breaking, the system will ignore the film breaking action and send a command to the traveling vehicle to collect the image of the next plant.

[0010] Further technology of the present invention:

[0011] Preferably, the specific method of step (1) is as follows:

[0012] 1) Collect a batch of peanut leaf images that meet the detection target. The camera is located on the guide rail or at a fixed height between the rows to collect images, and select the peanut plant images that meet the requirements for vertical shooting.

[0013] 2) The selected blades that meet the requirements are subjected to Gaussian filtering and image enhancement processing to make the image details clearer. The blade areas are marked with rectangular boxes, and then input into the Darknet-53 backbone network for training. The Gaussian filtering satisfies the following:

[0014]

[0015] Where (x,y) are the coordinates of any point inside the membrane, and σ is the standard deviation.

[0016] 3) The neural network will extract the interest features of the leaves. After training, a model file and a weight file will be obtained. The neural network outputs a set of loss parameters. At this time, in order to improve the accuracy and generalization of the neural network in recognizing targets, some peanut leaf images will be collected.

[0017] 4) After processing the second acquired image by flipping, occlusion, etc., it is input into the above neural network to update the model parameters. The Loss parameter is the error loss parameter, which can be used to evaluate the training effect. It includes the position error loss function Losspos, the confidence error loss function Lossconf, and the classification loss function Lossclass. Their relationship is as follows:

[0018] Loss total =Loss pos +Loss conf +Loss class (2)

[0019] Loss pos =λ coord ×CIoU_loss (3)

[0020]

[0021]

[0022] Where S 2 This indicates the number of grids into which the image is divided by the model. B represents the number of anchor boxes. and These represent whether a target exists in the j-th anchor frame within the i-th grid. and Let λ represent the actual confidence value and the predicted confidence value of the j-th anchor box in the i-th grid, respectively. coord ,λ obj ,λ noobj and λ class These are weighting coefficients used to balance different types of loss. and These represent the actual category information and the predicted category information of the j-th anchor box in the i-th grid, respectively.

[0023] Preferably, in step (3), for plants that meet the requirements for breaking the film, the three-dimensional coordinates of the plant are obtained according to the camera calibration parameters, the leaves are subjected to Canny edge detection, the center coordinates of the circle are determined by taking three points on the edge, the radius of the circle is calculated, the film breaking vehicle moves to the center position, the film breaking knife assembly adjusts the radius, and the ring cutting film breaking action is performed.

[0024] Preferably, in order to obtain the three-dimensional coordinates of the plant, the camera for acquiring the images is calibrated. The calibration method adopts the classic single-plane checkerboard camera calibration method, also known as the Zhang Zhengyou calibration method.

[0025] Preferably, the specific method of Zhang Zhengyou's calibration method is as follows:

[0026] 1) Print a standard chessboard template on a flat calibration board and take several pictures from different angles.

[0027] 2) Detect feature points in the image.

[0028] 3) Solve for the intrinsic and extrinsic parameters of the camera under ideal conditions, ignoring the effects of distortion, and substitute them into the maximum likelihood estimation method to improve accuracy.

[0029] 4) Solve for the radial distortion parameters using the least squares regression method, and eliminate the distortion using the inverse transformation.

[0030] 5) Optimize the estimation using the maximum likelihood estimation method, and input the distortion parameters to obtain the camera intrinsic parameters, namely the fusion of principal point coordinates, focal length, and pixel aspect ratio, as well as the radial distortion parameters.

[0031] Preferably, the method for obtaining the three-dimensional coordinates of the plant by camera calibration parameters is derived from the algorithm, that is, satisfying: x~K[R|t]X(6);

[0032] Where x is the camera coordinate; X is the real-world coordinate; K is the intrinsic parameter matrix; [R|t] is the extrinsic parameter matrix. K is the intrinsic parameter matrix, a 3×3 matrix composed of the camera's internal parameters, where f represents the focal length; s is the distortion parameter, (x0, y0)x0=αy0 are the center point coordinates, α is the aspect ratio parameter, which can be set to 1 by default, so x0=αy0. [R|t] is the extrinsic parameter matrix, R is the rotation matrix describing the camera's orientation, and t is the three-dimensional translation vector describing the camera's center position, i.e.:

[0033]

[0034] Preferably, the peanut plant image is acquired in real time by the film-breaking machine component. A bilateral filter is used to blur the entire image, reducing the appearance of soil particles while maintaining the color gradient of the leaves compared to their surroundings. Then, image binarization is performed to obtain a peanut leaf image with clear contrast. The bilateral filter satisfies the following:

[0035] Where K p p = (p is the image normalization factor) x ,p y ) represents the center pixel position, q = (q x ,q y ) represents the position of another pixel calculated by it, I q Let q be the pixel value of pixel q, and g be the Gaussian distribution kernel.

[0036] Compared with the prior art, the present invention has the following technical effects:

[0037] 1. The film-breaking process of this method largely eliminates the need for manual intervention. Simply calibrate the device and place it between rows to identify the seedlings that need film breaking, saving labor costs and improving production efficiency.

[0038] 2. This invention has high accuracy in identifying peanut seedlings. Even peanut seeds of the same variety and batch may have different growth rates due to differences in planting environment. Manual seed film breaking relies to some extent on production experience. This method uses appropriate processing algorithms for the identification environment and a model with high identification accuracy, which improves the robustness of the system. It can solve the problem once and for all in the production stage and greatly reduce the probability of false breaking.

[0039] 3. Traditional film-breaking devices mainly rely on mechanical structures. While conical film-breaking blades are easy to control, the breaking process is relatively violent and can easily damage the blades. This invention uses a circular rotating blade, relying on a positioning algorithm, resulting in a lighter and more compact design. This reduces mechanical and human errors, significantly improving film-breaking accuracy. This method performs a ring-cutting action after positioning the film, eliminating the need for puncture and grinding in traditional film-breaking machines. The cut film is also easier to recycle.

[0040] Other features and advantages of the present invention will be described in detail in the following specific embodiments. Attached Figure Description

[0041] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0042] Figure 1 Flowchart of the system executing the membrane rupture action after model deployment;

[0043] Figure 2 A schematic diagram of the training algorithm for a peanut plant identification model;

[0044] Figure 3 This is a schematic diagram of the YOLOv3 neural network structure, illustrating the neural network structure and the process of residual network computation;

[0045] Figure 4 Peanut leaves to be identified under the film;

[0046] Figure 5 To enable the system to identify peanut leaf selection frames that meet the requirements;

[0047] Figure 6 A schematic diagram illustrating edge detection and localization of the membrane rupture point;

[0048] Figure 7 This is a schematic diagram of peanut leaves after the membrane has been broken.

[0049] Figure 8 The structure of the automatic peanut seedling film breaking machine designed for this invention;

[0050] Figure 9 View of the drive motor of the film breaking machine;

[0051] Figure 10 This is a schematic diagram of the internal structure of the cam box.

[0052] The components are as follows: 1. Fixed terminal; 2. Cam housing; 3. Camera; 4. Connecting rod; 5. Motor board; 6. Guide rail; 7. Cutting blade; 8. Third motor; 9. Second motor; 10. Coupling; 11. Slider; 12. Nut seat; 13. Lead screw seat one; 14. Ball screw; 15. Lead screw seat two; 16. First motor; 17. Fourth motor; 18. Crankshaft; 19. Drive gear; 20. Driven gear; 21. Positioning ring. Detailed Implementation

[0053] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further explained below with reference to the specific accompanying drawings.

[0054] A method for breaking the film on a film-breaking machine component used in peanut cultivation, characterized by the following steps:

[0055] (1) Input the peanut leaf image into the YOLOv3 neural network model for training. After the model detects, it outputs the leaf detection result. The detection result is whether the current growth status of the peanut plant meets the requirements for breaking the film.

[0056] The training section first configures the YOLOv3 neural network environment, builds the YOLOv3 feature extraction network, and the data acquisition and training process is as follows: Figure 2 As shown, the specific process is as follows:

[0057] 1) Collect a batch of peanut leaf images that meet the detection target. The camera is located on the guide rail or at a fixed height between the rows to collect images, and select the vertically shot peanut plant images that meet the requirements.

[0058] 2) The selected blades that meet the requirements are subjected to Gaussian filtering and image enhancement processing to make the image details clearer. The blade areas are marked with rectangular boxes, and then input into the Darknet-53 backbone network for training. The Gaussian filtering satisfies the following:

[0059]

[0060] Where (x,y) are the coordinates of any point inside the membrane, and σ is the standard deviation;

[0061] 3) Neural networks extract interest features from the leaves, such as... Figure 3 As shown, the extraction process involves convolution and pooling. The neural network detects the color gradient of the region of interest, and multiple channels are used for weight updates. After training, a model file and a weight file will be obtained. The neural network outputs a set of loss parameters. To improve the accuracy and generalization of the neural network in recognizing targets, some peanut leaf images are collected.

[0062] 4) After processing the second acquired image by flipping, occlusion, etc., it is input into the above neural network to update the model parameters. The Loss parameter is the error loss parameter, which can be used to evaluate the training effect. It includes the position error loss function Losspos, the confidence error loss function Lossconf, and the classification loss function Lossclass. Their relationship is as follows:

[0063] Loss total =Loss pos +Loss conf +Loss class (2)

[0064] Loss pos =λ coord ×CIoU_loss (3)

[0065]

[0066]

[0067] in:

[0068] S 2 This indicates the number of grids into which the image is divided by the model;

[0069] B indicates the number of anchor frames;

[0070] and These indicate whether a target exists in the j-th anchor frame within the i-th grid.

[0071] and Let represent the actual confidence value and the predicted confidence value of the j-th anchor box in the i-th grid, respectively;

[0072] λ coord ,λ obj ,λ noobj and λ class These are weighting coefficients used to balance different types of loss;

[0073] and These represent the actual category information and the predicted category information of the j-th anchor box in the i-th grid, respectively.

[0074] (2) Figure 1 As shown, after the training is completed and the model is deployed, the trained peanut leaf recognition model will obtain the corresponding model file. At this time, the training results are converted into a format and quantized and optimized before being deployed to the embedded MCU system. The embedded MCU system is connected to the film breaking machine component, which is mounted on the inter-row traveling vehicle.

[0075] (3) The film breaking machine component follows the traveling vehicle to collect images in real time. Figure 4 The acquired image to be detected is used as the basis for the neural network to output a judgment result based on the leaf image. Figure 5 As shown, for plants that meet the requirements for breaking the film, the system outputs a film breaking command, acquires the three-dimensional coordinates, and executes the film breaking action; for leaves that do not meet the requirements for breaking the film, the system ignores the film breaking action and sends a command to the traveling vehicle to acquire images of the next plant.

[0076] Obtaining the three-dimensional coordinates of peanut plants relies on the calibration calculation of camera parameters. This process requires combining the camera's intrinsic and extrinsic parameters, distortion correction algorithms, and using two-dimensional images to calculate the object's three-dimensional coordinates. The calibration method can employ the classic single-plane checkerboard camera calibration method, also known as the Zhang Zhengyou calibration method. The basic calibration process is as follows:

[0077] (1) Print a standard chessboard template on a flat calibration board and take several pictures from different angles.

[0078] (2) Detect feature points in the image.

[0079] (3) Solve for the intrinsic and extrinsic parameters of the camera under ideal conditions without regard to distortion, and substitute them into the maximum likelihood estimation method to improve accuracy.

[0080] (4) The radial distortion parameters are solved by regression using the least squares method, and the distortion is eliminated by inverse transformation.

[0081] (5) Continue to use the maximum likelihood estimation method to optimize the estimation, and substitute the distortion parameters to find the six degrees of freedom of the camera intrinsic parameters, namely the fusion of principal point coordinates, focal length and pixel aspect ratio and radial distortion parameters.

[0082] The camera calibration 3D conversion method is derived from an algorithm, which satisfies: x~K[R|t]X(6)

[0083] Where x is the coordinate in the camera; X is the real-world coordinate; K is the intrinsic parameter matrix; [R|t] is the extrinsic parameter matrix. K is the intrinsic parameter matrix, which is a 3×3 matrix composed of the camera's intrinsic parameters.

[0084] The peanut plant image acquisition machine captures images of peanut plants in real time, but it is easily affected by noise during edge detection. Furthermore, the peanut planting environment is complex, necessitating image filtering. Considering the characteristics of the peanut planting environment, soil particles are the most likely factor to affect recognition. This invention employs bilateral filtering to blur the entire image, reducing the appearance of soil particles while maintaining the color gradient of the leaves compared to their surroundings. Then, image binarization processing yields a peanut leaf image with clear contrast. The bilateral filtering satisfies the following:

[0085]

[0086] Where K p p = (p is the image normalization factor) x ,p y ) represents the center pixel position, q = (q x ,q y ) represents the position of another pixel calculated by it, I q Let q be the pixel value of pixel q, and g be the Gaussian distribution kernel.

[0087] like Figure 6 As shown, after image preprocessing, the image begins Canny edge detection. First, Gaussian filtering is applied to the image to further reduce noise and smooth the image. Then, the gradient of the pixels in the image is calculated to obtain the x and y dimensions of the image. This dimension and the image scale are returned to the calibration algorithm to obtain the world coordinates of the peanut seedling mulch. After recognition, the system will obtain the judgment information of whether the peanut seedling meets the requirements for mulch breaking and the calculated seedling coordinates. The mulch breaking structure linkage used in this invention is a cam structure. Therefore, the system needs to send the center coordinates to the mulch breaking vehicle, and the vehicle-mounted mulch breaking machine moves to the center position and then performs the mulch breaking action.

[0088] To ensure the scientific and safe nature of the film breaking process, this invention no longer uses the traditional conical film breaking blade that is prone to damaging the blades. Instead, it uses a motor-driven rotating blade film breaking blade. This type of film breaking blade is sharper and more efficient than the conical one, and it is also smaller in size, which can reduce mechanical errors caused by mechanical design.

[0089] To ensure safer membrane breaking and scientific cultivation of peanut seedlings, this invention no longer uses the actual world coordinates of the peanut leaf as the membrane breaking coordinates. Instead, it uses three points on the two-dimensional outline to draw a complete circle, such as... Figure 7 As shown, after the circular radius is increased by one centimeter, the system calculates the circular radius, and the rupture knife obtains the circular radius and locates the center of the circle before performing the action.

[0090] To achieve the film-breaking method of the present invention, a film-breaking machine component for peanut cultivation is provided, such as... Figure 8-10 The system includes a cam housing 2 mounted on an inter-row traveling vehicle, a camera 3 mounted on the cam housing 2, and the cam housing 2 is fixed to the inter-row traveling vehicle via a fixed terminal 1. The cam housing 2 is provided with a positioning ring 21, and the bottom of the connecting rod 4 extends through the positioning ring 21 and extends out of the housing to movably connect to a motor plate 5. A first motor 16 is provided below the motor plate 5, and a drive gear 19 is provided on the first motor 16. A driven gear 20 is provided on the connecting rod 4, and the drive gear 19 and the driven gear 20 mesh.

[0091] The cam housing 2 is equipped with a crankshaft 18 driven by a fourth motor 17, and the connecting rod 4 is connected to the crankshaft 18.

[0092] Below the motor plate 5, there is a guide rail 6, on which a slider 11 slides. Below the motor plate 5, there is also a second motor 9, a first lead screw seat 13, and a second lead screw seat 15. A ball screw 14 is inserted into the first lead screw seat 13 and the second lead screw seat 15. The second motor 9 is connected to the ball screw 14 through a coupling 10. A nut seat 12 is connected to the slider 11, and the nut seat 12 is also connected to the nut of the ball screw 14. A third motor 8 is provided on the nut seat 12, and a cutting blade 7 is connected to the output shaft of the third motor 8.

[0093] The first motor 16, the second motor 9, the third motor 8, and the fourth motor 17 mentioned above are all driven by a driver. The driver receives the PID parameters calculated by the embedded MCU system based on the position and converts them into PWM signals to control the rotation of the motor.

[0094] Before breaking the film, the two fixed terminals 1 are first fixed along with the inter-row traveling vehicle, and the camera 3 is fixed at the bottom of the connecting rod 4. Before detecting the blade, the cam structure is in the raised state. When the film breaking action begins, the embedded MCU system sends the film breaking radius and positions the cam structure in the appropriate position. After the motor adjusts the rotation radius, it feeds back to the embedded MCU system and begins to execute the ring cutting film breaking action.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for breaking the film of a film-breaking machine component used in peanut cultivation, characterized in that, The steps are as follows: (1) Input peanut leaf images into the YOLOv3 neural network model for training. After the model detects, it outputs the leaf detection results. The detection results are whether the current growth status of the peanut plant meets the requirements for breaking the film. (2) After the training is completed and the model is deployed, the peanut leaf recognition model will obtain the corresponding model file. At this time, the training results are converted into a format and then optimized by quantization before being deployed to the embedded MCU system. The embedded MCU system is connected to the film breaking machine component, which is mounted on the inter-row traveling vehicle. (3) The film breaking machine component follows the traveling vehicle to collect images in real time. The neural network will output the judgment result based on the leaf image of the collected images. For plants that meet the requirements for film breaking, the system outputs the film breaking command, obtains the three-dimensional coordinate process, and executes the film breaking action. For leaves that do not meet the requirements for film breaking, the system will ignore the film breaking action and send a command to the traveling vehicle to collect images of the next plant. In step (3), for plants that meet the requirements for breaking the film, the three-dimensional coordinates of the plants are obtained according to the camera calibration parameters. Canny edge detection is performed on the leaves, and the center coordinates of the circle are determined by taking three points on the edge. The radius of the circle is calculated. The film breaking vehicle moves to the center position, the film breaking knife assembly adjusts the radius, and the ring cutting film breaking action is performed. In order to obtain the three-dimensional coordinates of the plant, the parameters of the camera used to acquire the images were calibrated. The calibration method adopted was the classic single-plane checkerboard camera calibration method, also known as the Zhang Zhengyou calibration method. The specific method of Zhang Zhengyou's calibration method is as follows: 1) Print a standard chessboard template on a flat calibration board and take several pictures from different angles; 2) Detect feature points in the image; 3) Solve for the intrinsic and extrinsic parameters of the camera under ideal conditions, ignoring the effects of distortion, and substitute them into the maximum likelihood estimation method to improve accuracy; 4) Solve for the radial distortion parameters using the least squares regression method, and eliminate the distortion using the inverse transformation; 5) Optimize the estimation using the maximum likelihood estimation method, and input the distortion parameters to obtain the camera intrinsic parameters, namely the fusion of principal point coordinates, focal length and pixel aspect ratio, and radial distortion parameters. The method for obtaining the three-dimensional coordinates of the plant by camera calibration parameters is derived from the algorithm, which satisfies: x~K[R|t]X(6); Where x is the coordinate in the camera; X is the real-world coordinate; K is the intrinsic parameter matrix; [R|t] is the extrinsic parameter matrix. K is the intrinsic parameter matrix, a 3×3 matrix composed of the camera's internal parameters, where f represents the focal length; s is the distortion parameter, (x0, y0)x0=αy0 are the coordinates of the center point, and α is the aspect ratio parameter; [R|t] is the extrinsic parameter matrix, where R is the rotation matrix describing the camera's orientation, and t is the three-dimensional translation vector describing the camera's center position, i.e.: In step (3), the film-breaking machine component acquires peanut plant images in real time, and uses bilateral filtering to blur the entire image, reducing the soil particle effect while ensuring the difference in leaf color gradient with the surrounding area. Then, image binarization processing is performed to obtain peanut leaf images with obvious contrast. The bilateral filtering satisfies the following: Where K p p = (p is the image normalization factor) x ,p y ) represents the center pixel position, q = (q x ,q y ) represents the position of another pixel calculated by it, I q Let q be the pixel value of pixel q, and g be the Gaussian distribution kernel.

2. The film-breaking method for a film-breaking machine assembly used in peanut cultivation according to claim 1, characterized in that: The specific method of step (1) is as follows: 1) Collect a batch of peanut leaf images that meet the detection target. The camera is located on the guide rail or at a fixed height between the rows to collect images, and select the vertically shot peanut plant images that meet the requirements. 2) The selected blades that meet the requirements are subjected to Gaussian filtering and image enhancement processing to make the image details clearer. The blade areas are marked with rectangular boxes, and then input into the Darknet-53 backbone network for training. The Gaussian filtering satisfies the following: Where (x,y) are the coordinates of any point inside the membrane, and σ is the standard deviation; 3) The neural network will extract the interest features of the leaves. After training, a model file and a weight file will be obtained. The neural network outputs a set of loss parameters. At this time, in order to improve the accuracy and generalization of the neural network in recognizing targets, some peanut leaf images will be collected. 4) After flipping and occluding the second acquired image, input it into the neural network to update the model parameters. The Loss parameters are the error loss parameters, which can be used to evaluate the training effect. They include the location error loss function Losspos, the confidence error loss function Lossconf, and the classification loss function Lossclass. Their relationship is as follows: Loss total =Loss pos +Loss conf +Loss class (2) Loss pos =λ coord ×CIoU_loss (3) in: S 2 This indicates the number of grids into which the image is divided by the model; B indicates the number of anchor frames; and These indicate whether a target exists in the j-th anchor frame within the i-th grid. and Let represent the actual confidence value and the predicted confidence value of the j-th anchor box in the i-th grid, respectively; λ coord ,λ obj ,λ noobj and λ class These are weighting coefficients used to balance different types of loss; and These represent the actual category information and the predicted category information of the j-th anchor box in the i-th grid, respectively.

Images