A soil removal method for potato harvester based on E-NET

Abstract

The present invention relates to the technical field of potato harvesting, and in particular provides an E-NET-based clod removal method for a potato harvester. The method, based on the E-NET algorithm, proposes an E-NET model with a simple structure and good real-time performance. The model has good real-time performance, low computational complexity, and low memory usage, and can meet the requirements of potato and debris separation scenarios. Deep learning and machine vision algorithms are used to recognize and locate clods, thereby determining target striking points for the clods, improving target striking accuracy, and breaking large clods into small clods. The small clods pass through gaps between conveyor chain rollers and fall to the ground, thereby achieving clod removal. The method can accurately remove clods from potatoes and is suitable for continuous operation.

Description

Technical Field

[0001] The present application relates to the technical field of potato harvesting, and in particular to a soil clod removal method for a potato harvester based on E-NET. Background Art

[0002] Potatoes are the world's most important food crop after wheat, rice, and corn. China currently produces the most potatoes in the world. However, the widespread cultivation of potatoes makes harvesting them a challenge. Currently, removing clods from the potato stream during harvest remains a challenge, requiring manual labor. This process is both inefficient and labor-intensive.

[0003] The impact separation device designed by Feller R et al. uses the difference in the rebound trajectory of the steel roller to achieve separation. Although the separation success rate is high, the falling height needs to be strictly controlled, which limits the processing speed, and the complex mechanical structure increases maintenance costs.

[0004] Liu Wendong and others developed a robotic picking system based on visual recognition. It identifies soil clods through external features, which provides ideas for automated cleaning. However, it relies on high-precision sensors and computer processing, has high hardware costs, and its dynamic response speed cannot meet the needs of continuous operations.

[0005] The machine vision cleaning and control system developed by Wang Zedong of Heilongjiang Bayi Agricultural University has a high degree of system integration and a long debugging cycle, making it difficult to popularize in small and medium-sized agricultural machinery. Summary of the Invention

[0006] In order to solve the above technical problems, the present application provides a method for removing soil blocks for a potato harvester based on E-NET, the method comprising the following steps:

[0007] S1, real-time image acquisition during the impurity removal process;

[0008] S2, preprocessing the image to reduce image noise;

[0009] S3, using the E-NET algorithm to perform instance segmentation on the preprocessed image;

[0010] S4, outputs the mask layer image after instance segmentation; the instance and the background are marked with different colors;

[0011] S5, determining the target hitting point according to the mask layer image;

[0012] S6, determining the coordinates of the target impact point;

[0013] S7, executing the striking action according to the coordinates of the target striking point.

[0014] In some embodiments of the present application, in step S5, determining the target hitting point according to the mask layer image includes the following steps:

[0015] S51, determining a strikeable trajectory based on the strike execution component and the motion path of the conveyor belt;

[0016] S52, determining the intersection of the strikeable trajectory and the soil block mask according to the strikeable trajectory and the mask layer image after instance segmentation;

[0017] S53: Determine the target striking point according to the intersection of the strikeable trajectory and the soil clod mask.

[0018] In some embodiments of the present application, in step S53, the center point of the two intersection points is configured as the target hitting point.

[0019] In some embodiments of the present application, in step S7, a striking device is used to perform the striking action; the striking device includes a conveyor belt, a striking execution component and a camera; potatoes and clods to be identified are placed on the conveyor belt; the camera and the striking execution component are respectively arranged above the conveyor belt along the conveying direction of the potatoes.

[0020] In some embodiments of the present application, the striking execution component includes two groups of cylinder assemblies arranged side by side, and each group of cylinder assemblies includes a plurality of cylinders arranged side by side.

[0021] In some embodiments of the present application, in step S6, it is assumed that the coordinate system of the camera detecting the object position is ( ), the coordinate system of the hitting execution component is ( ), the relationship between the coordinate system of the camera detecting the object position and the coordinate system of the hitting actuator performing the action is shown in the following formula:

[0022]

[0023] Where: X represents a 3×3 homogeneous transformation matrix, and its specific structure is as follows;

[0024]

[0025] and is the rotation component in the X-axis direction; is the translation in the X-axis direction; and is the rotation component in the Y-axis direction; is the translation in the Y-axis direction;

[0026] The calculation process of X is as follows:

[0027] Configure 9 standard targets in the striking area and place the targets in the coordinate system ( ) is configured as ( , ), place the target in the coordinate system ( ) is configured as ( , )

[0028] The coordinate relationship of each target point is as follows:

[0029] .

[0030] In some embodiments of the present application, in step S2, preprocessing the image includes cropping and filtering;

[0031] The cropping is used to adjust the size and shape of the image;

[0032] The filtering adopts bilateral filtering.

[0033] In some embodiments of the present application, the E-NET algorithm includes a linear deformable convolution layer.

[0034] In some embodiments of the present application, the E-NET algorithm sets three detection heads, namely, a P3 / 8 detection head, a P4 / 16 detection head, and a P5 / 32 detection head;

[0035] The detection feature map size corresponding to the P3 / 8 detection head is 80*80, which is used to detect targets larger than 8*8.

[0036] The detection feature map size of the P4 / 16 detection head is 40*40, which is used to detect targets larger than 16*16.

[0037] The detection feature map size corresponding to the P5 / 32 detection head is 20*20, which is used to detect targets larger than 32*32.

[0038] In some embodiments of the present application, the detection head includes a bounding box prediction branch and a classification prediction branch;

[0039] The bounding box prediction branch includes two convolutional layers, a two-dimensional convolution and a bounding box loss module;

[0040] The classification prediction branch includes two depth-wise separable convolutions, two convolutional layers, a two-dimensional convolution, and a classification loss module.

[0041] Compared with the prior art, the present invention has the following advantages and beneficial effects: the E-NET-based clod removal method for potato harvesters of the present application proposes an E-NET model with simple structure and good real-time performance based on the E-NET algorithm. The model has good real-time performance, small amount of calculation, and small memory occupation, and can meet the requirements of potato and offal separation scenarios; deep learning and machine vision algorithms are used to realize the recognition and positioning of clods, and then determine the target hitting point of the clods, improve the accuracy of the target hitting, break large clods into small clods, and the small clods fall to the ground through the gaps between the conveyor chain rollers to realize the removal of clods. This method can accurately remove clods in potatoes and is suitable for continuous operation.

[0042] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which constitute part of this document, are intended to provide a further understanding of this document. The exemplary embodiments and descriptions herein are intended to explain this document and do not constitute an improper limitation on this document. In the accompanying drawings:

[0044] Figure 1 This is a principle block diagram of a soil block removal method for a potato harvester based on E-NET provided by an exemplary embodiment of the present application;

[0045] Figure 2 is a perspective view of a striking device provided by an exemplary embodiment of the present application;

[0046] Figure 3 is a side view of a potato harvester provided by an exemplary embodiment of the present application;

[0047] Figure 4 This is a comparison diagram before and after filtering provided by an exemplary embodiment of the present application;

[0048] Figure 5 is a process diagram for determining a target hitting point provided by an exemplary embodiment of the present application;

[0049] Figure 6 It is an E-NET algorithm model diagram provided by an exemplary embodiment of the present application;

[0050] Figure 7 This is a diagram of a linear deformable convolutional layer structure model in the E-NET algorithm provided by an exemplary embodiment of the present application;

[0051] Figure 8 This is a diagram of the convolutional layer structure model in the E-NET algorithm provided by an exemplary embodiment of the present application;

[0052] Figure 9This is a diagram of the residual block structure model in the E-NET algorithm provided by an exemplary embodiment of the present application;

[0053] Figure 10 This is a diagram of a detection head structure model in an E-NET algorithm provided by an exemplary embodiment of the present application;

[0054] Figure 11 This is a Yolo v11n algorithm test effect diagram provided by an exemplary embodiment of the present application;

[0055] Figure 12 This is an E-NET algorithm test effect diagram provided by an exemplary embodiment of the present application.

[0056] In the picture:

[0057] 1. Camera; 2. Light source; 3. Conveyor belt; 4. Striking execution component. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other in any way.

[0059] Potatoes are the world's most important food crop after wheat, rice, and corn. China currently produces the most potatoes in the world. However, the widespread cultivation of potatoes makes harvesting them a challenge. Currently, removing clods from the potato stream during harvest remains a challenge, requiring manual labor. This process is both inefficient and labor-intensive.

[0060] The impact separation device designed by Feller R et al. uses the difference in the rebound trajectory of the steel roller to achieve separation. Although the separation success rate is high, the falling height needs to be strictly controlled, which limits the processing speed, and the complex mechanical structure increases maintenance costs.

[0061] Liu Wendong and others developed a robotic picking system based on visual recognition. It identifies soil clods through external features, which provides ideas for automated cleaning. However, it relies on high-precision sensors and computer processing, has high hardware costs, and its dynamic response speed cannot meet the needs of continuous operations.

[0062] The machine vision cleaning and control system developed by Wang Zedong of Heilongjiang Bayi Agricultural University has a high degree of system integration and a long debugging cycle, making it difficult to popularize in small and medium-sized agricultural machinery.

[0063] Based on this, the present application exemplarily provides an E-NET-based soil clod removal method for a potato harvester. Based on the E-NET algorithm, an E-NET model with a simple structure and good real-time performance is proposed. The model has good real-time performance, small computational complexity, and small memory usage, and can meet the requirements of potato and offal separation scenarios. Deep learning and machine vision algorithms are used to realize the recognition and positioning of soil clods, and then determine the target hitting point of the soil clods, improve the accuracy of the target hitting, and break large soil clods into small soil clods. The small soil clods pass through the gaps between the conveyor chain rollers and fall to the ground, thereby realizing the removal of soil clods. This method can accurately remove soil clods in potatoes and is suitable for continuous operations.

[0064] An exemplary embodiment of the present application provides a soil block removal method for a potato harvester based on E-NET, such as Figure 1 As shown, the method includes the following steps:

[0065] S1, real-time image acquisition during the impurity removal process;

[0066] S2, preprocessing the image to reduce image noise;

[0067] S3, using the E-NET algorithm to perform instance segmentation on the preprocessed image;

[0068] S4, outputs the mask layer image after instance segmentation; the instance and the background are marked with different colors;

[0069] S5, determining the target hitting point according to the mask layer image;

[0070] S6, determining the coordinates of the target impact point;

[0071] S7, executing the striking action according to the coordinates of the target striking point.

[0072] This application uses a striking device to implement and execute the striking action; Figure 2 and 3 As shown, after the harvested potatoes are initially screened to remove small clods of soil, the remaining potatoes and larger clods of soil are transported to a beating device, which includes a conveyor belt 3, a beating execution component 4, a light source 2, a bracket, and a camera 1; the bracket is mounted above the conveyor belt 3, and the camera 1 and the light source 2 are both arranged on the bracket. Preferably, the camera is a Hikvision industrial camera; the light source 2 is preferably a bar light source, and the symmetrical installation of the two bar light sources can not only provide sufficient brightness for the system, but also prevent the problem of poor quality of the collected pictures caused by light source reflection; in particular, the outside of the bracket is equipped with a light shield to avoid interference from external light sources. Potatoes and clods to be identified are placed on the conveyor belt 3; the beating execution component 4 is arranged above the conveyor belt 3, and the cylinder assembly is arranged downstream of the bracket along the conveying direction of the potatoes.

[0073] Preferably, the striking actuator 4 comprises two side-by-side cylinder assemblies, each comprising a plurality of side-by-side cylinders; the cylinders of the two cylinder assemblies are arranged alternately. For example, the first cylinder assembly comprises seven cylinders, and the second cylinder assembly comprises six cylinders. The cylinders of the first and second cylinder assemblies are staggered in a direction perpendicular to the conveyor belt conveyor direction, such that the spacing between striking points is 3.5 cm. While the spacing between adjacent conveyor chain rollers in a potato harvester is 5 cm, the crushed small soil clods can fall into the gaps between the next conveyor chain, thereby removing the soil clods.

[0074] In step S1, a camera is used to capture images.

[0075] In step S2, the image is pre-processed, including cropping and filtering. Cropping is used to adjust the size and shape of the image. Preferably, the size of the cropped image is 640*640. Bilateral filtering is used in this step. It is a nonlinear filtering method that combines the spatial proximity and pixel value similarity of the image, while considering spatial information and grayscale similarity to achieve the purpose of edge-preserving denoising. This method is simple, non-iterative, and localized, and is particularly suitable for processing edge information in images.

[0076] The formula for bilateral filtering can be expressed as:

[0077]

[0078] in: is the pixel value after filtering;

[0079] is the pixel value of the original image;

[0080] is the normalized weight, used to normalize the filtering results;

[0081] represents the neighborhood window of the filter;

[0082] It is a spatial domain kernel function used to calculate the spatial distance weight between pixels;

[0083] It is a pixel value domain kernel function, which is used to calculate the similarity weight between pixel values.

[0084] Because excessive filtering will lose the original texture, color and other features of the target, we usually only perform light filtering in actual operation. In this method, we set the value of d to 3; the value of sigmaColor to 20; and the value of sigmaSpace to 50. If the value of sigmaSpace is large, it means that pixels with similar colors that are farther away will affect each other, so that colors that are similar enough in a larger area will obtain the same color. The image comparison before and after filtering is as follows: Figure 4 shown.

[0085] S3, use the E-NET algorithm to perform instance segmentation on the preprocessed image, such as Figure 5 As shown in the segmented mask image, white represents soil blocks and black represents background. The E-NET algorithm is improved based on the Yolo v11n algorithm to form the E-NET algorithm. Its network model structure is as follows Figures 6 to 10 As shown in the figure, the E-NET algorithm includes input, convolutional layer, linear deformable convolutional layer, residual block, and detection head modules. The convolutional layer includes two-dimensional convolution, batch normalization, and activation function; the residual block includes convolutional layer, channel splitting, residual bottleneck module, and feature fusion; the residual bottleneck module includes two convolutional layers.

[0086] The E-NET model is based on yolo v11n. E-NET proposes a simple and practical structure for the needs of soil clod impact scenarios. Due to the different shapes of soil clods, standard convolution samples in a local window of fixed shape and size, which is difficult to dynamically adapt to the shapes of different objects. Although deformable convolution allows flexible sampling positions, the number of its parameters grows quadratically with the size of the convolution kernel, and the computational efficiency is low. Therefore, this application introduces a linear deformable convolution layer to adapt to soil clods of different shapes and sizes. The linear deformable convolution layer includes two-dimensional convolution, point-by-point convolution, and linear deformable convolution.

[0087] Linear deformable convolution proposes an algorithm to generate the initial sampling coordinates of convolution kernels of any size and adjusts the sampling shape through an offset so that it can adapt to changes in the target shape. At the same time, linear deformable convolution allows the convolution kernel to have any number of parameters, such as 1, 2, 3, 4, 5, 6, 7, etc., thus providing more flexibility for network design; at the same time, linear deformable convolution changes the growth trend of the number of parameters from quadratic to linear, thereby reducing the requirements for the hardware environment.

[0088] For input X, the implementation process of linear deformable convolution is as follows: First, generate initial sampling coordinates: Based on the size of the convolution kernel, use an algorithm to generate initial sampling coordinates. These coordinates can be any shape, such as a triangle, rectangle, or diamond. Second, calculate the offset: The offset is learned through the convolution operation and added to the initial sampling coordinates to generate new sampling coordinates, thereby adjusting the sampling shape. Third, extract features: By interpolating and resampling the feature map, the features corresponding to the new sampling coordinates are obtained, and the corresponding convolution operation is used to extract features.

[0089] In the E-NET algorithm, the input image first enters the convolutional layer to extract preliminary features. The extracted features are then sequentially passed through a combination of linear deformable convolutional layers and residual blocks. The linear deformable convolutional layers further extract flexible and diverse features, while the residual blocks help the network better train and learn deep features. After processing through three sets of linear deformable convolutional layers and residual blocks, the images are input to the corresponding detection heads at the 4th, 6th, and 8th layers, respectively. The detection heads make predictions related to object detection based on these features and output information such as the object's category and location.

[0090] The E-NET algorithm sets three detection heads, namely the P3 / 8 detection head, the P4 / 16 detection head, and the P5 / 32 detection head. The detection feature map size corresponding to the P3 / 8 detection head is 80*80, which is used to detect targets larger than 8*8. The detection feature map size corresponding to the P4 / 16 detection head is 40*40, which is used to detect targets larger than 16*16. The detection feature map size corresponding to the P5 / 32 detection head is 20*20, which is used to detect targets larger than 32*32.

[0091] like Figure 10 As shown in the figure, the detection head consists of a bounding box prediction branch and a classification prediction branch. The bounding box prediction branch consists of two convolutional layers, a two-dimensional convolution, and a bounding box loss module. The features output by the detection head are first processed by the convolutional layer to extract preliminary features, and then further processed by subsequent convolutional layers and two-dimensional convolutions to continuously refine the features. Ultimately, they are used to calculate the bounding box loss and optimize the target position prediction. The classification prediction branch consists of two depthwise separable convolutions, two convolutional layers, a two-dimensional convolution, and a classification loss module. The detection head features enter the depthwise separable convolution, using its efficient feature extraction method. The features are further processed in combination with the convolutional layers. After multiple operations, the results are output through two-dimensional convolution, and the classification loss is calculated to optimize the target category prediction.

[0092] On the same device, we configured the Python 3.9 environment and the GTX 1650 (4G) graphics card to test the same test set (200 images). The experimental results are as follows: Figure 11 、 Figure 12 And shown in Table 1.

[0093] Table 1 Comparative test data of Yolo v11n algorithm and E-NET algorithm

[0094]

[0095] from Figure 11 and Figure 12 It can be seen that the E-NET model is slightly worse than Yolo v11n in segmentation accuracy, but the E-NET model is generally better than Yolo v11n in confidence, and the actual effect of the E-NET model can already meet the needs of soil separation scenarios.

[0096] As shown in Table 1, the E-NET model achieves a 0.001 increase in precision over Yolo v11n, maintains the same recall as Yolo v11n, and decreases the mean average precision by 0.01. Due to the redesigned network structure and the introduction of linear variable linear convolution, the E-NET model has significant advantages in model size and real-time performance. Compared to Yolo v11n, the E-NET model size has been reduced from 6.0 MB to 3.6 MB, and the frame rate has been increased from 7 fps to 13 fps.

[0097] In step S5, the target hitting point is determined according to the mask layer image, and the Figure 5 , including the following steps:

[0098] S51, determine the strikeable trajectory based on the striking execution component and the movement path of the conveyor belt; illustratively, 13 cylinders are set in the present application, and the cylinders are arranged in a direction perpendicular to the conveying direction of the conveyor belt. Each cylinder performs piston movement in the vertical direction to strike the soil blocks. When the conveyor belt moves, each cylinder will form a strikeable trajectory in the striking area, totaling 13 strikeable trajectories.

[0099] S52 , determining the intersections of the strikeable track and the clod mask according to the strikeable track and the mask layer image after instance segmentation; along the extension direction of the strikeable track, each strikeable track has two intersections with the clod mask.

[0100] S53: Determine the target striking point according to the intersection of the strikeable trajectory and the soil clod mask.

[0101] Preferably, in order to accurately hit the soil clods and break them into small clods of relatively uniform size, the center point of the two intersections of each hittable trajectory and the soil clod mask is configured as the target hitting point.

[0102] S6, determine the coordinates of the target hitting point; in this application, the coordinates of the hitting point are obtained by converting the pixel coordinates into the actual coordinates. Specifically, the coordinates of the hitting point are obtained by multiplying the pixel coordinates by the conversion matrix. Assume that the pixel coordinate system is the coordinate system of the camera detecting the object position. The coordinate system of the striking point is , the relationship between the coordinate system of the camera detecting the object position and the coordinate system of the hitting execution component performing the action is shown in the following formula:

[0103]

[0104] Where: X represents a 3×3 homogeneous transformation matrix, and its specific structure is as follows;

[0105]

[0106] and is the rotation component in the X-axis direction; is the translation in the X-axis direction; and is the rotation component in the Y-axis direction; is the translation in the Y-axis direction;

[0107] The calculation process of X is as follows:

[0108] Configure 9 standard targets in the striking area and place the targets in the coordinate system. The coordinate configuration in is ( , ), place the target in the coordinate system The coordinate configuration in is ( , )

[0109] The coordinate relationship of each target point is as follows:

[0110]

[0111] By calculating the coordinate relationship of the 9 target points and solving the simultaneous equations, X can be solved and the solved X can be substituted into the formula , you can get the coordinates of the hitting point.

[0112] S7: Execute the striking action based on the coordinates of the target striking point. Exemplarily, after determining the coordinates of the target striking point, the control system transmits the coordinates to the lower computer. The lower computer receives the striking command. When the soil clump moves along the conveyor belt to the bottom of the striking actuator, the striking actuator executes the striking action, breaking the large soil clump into small pieces. Exemplarily, in the coordinates of the striking point, C represents the cylinder number, with a value range of 01-13, corresponding to the 13 cylinders of the striking actuator, each equipped with a solenoid valve. E represents the number of pulses from the photoelectric encoder, with a value range of 0-9999. For example, (02, 1000) means that when the microcontroller receives 1000 pulses from the photoelectric encoder, it outputs a high level to the solenoid valve of cylinder No. 2. The solenoid valve controls the opening and closing of the cylinder's air inlet and outlet, causing the cylinder to strike.

[0113] In this application, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such article or device. In the absence of further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the article or device comprising the element.

[0114] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0115] Obviously, those skilled in the art may make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if such changes and modifications of the present application fall within the scope of the claims of the present application and their equivalents, the present application is intended to include such changes and modifications.

Claims

1. A soil block removal method for potato harvester based on E-NET, characterized in that: The method comprises the following steps: S1, real-time image acquisition during the impurity removal process; S2, preprocessing the image to reduce image noise; S3, using the E-NET algorithm to perform instance segmentation on the preprocessed image; S4, outputs the mask layer image after instance segmentation; the instance and the background are marked with different colors; S5, determining the target hitting point according to the mask layer image; Wherein, step S5 includes the following steps: S51, determining a strikeable trajectory based on the strike execution component and the motion path of the conveyor belt; S52, determining the intersection of the strikeable trajectory and the soil block mask according to the strikeable trajectory and the mask layer image after instance segmentation; S53, according to the intersection of the strikeable trajectory and the soil block mask, configuring the center point of two intersection points on the same strikeable trajectory as the target strike point; S6, determining the coordinates of the target impact point; S7, executing the striking action according to the coordinates of the target striking point.

2. The method for removing soil blocks for potato harvesters based on E-NET according to claim 1, characterized in that: In step S7, a striking device is used to perform the striking action; the striking device includes a conveyor belt, a striking execution component and a camera; potatoes and clods to be identified are placed on the conveyor belt; the camera and the striking execution component are respectively arranged above the conveyor belt along the conveying direction of the potatoes.

3. The soil block removal method for potato harvester based on E-NET according to claim 2, characterized in that: The striking execution component includes two groups of cylinder assemblies arranged side by side, and each group of cylinder assemblies includes a plurality of cylinders arranged side by side.

4. The method for removing soil blocks for a potato harvester based on E-NET according to claim 1, characterized in that: In step S6, it is assumed that the coordinate system of the camera detecting the object position is ( ), the coordinate system of the hitting execution component is ( ), the relationship between the coordinate system of the camera detecting the object position and the coordinate system of the hitting actuator performing the action is shown in the following formula: Where: X represents a 3×3 homogeneous transformation matrix, and its specific structure is as follows; and is the rotation component in the X-axis direction; is the translation in the X-axis direction; and is the rotation component in the Y-axis direction; is the translation in the Y-axis direction; The calculation process of X is as follows: Configure 9 standard targets in the striking area and place the targets in the coordinate system ( ) is configured as ( , ), place the target in the coordinate system ( ) is configured as ( , ) The coordinate relationship of each target point is as follows: 。 5. The method for removing soil blocks for a potato harvester based on E-NET according to claim 1, characterized in that: In step S2, pre-processing the image includes cropping and filtering; The cropping is used to adjust the size and shape of the image; The filtering adopts bilateral filtering.

6. The soil block removal method for potato harvester based on E-NET according to claim 1, characterized in that: The E-NET algorithm includes linear deformable convolutional layers.

7. The soil block removal method for potato harvester based on E-NET according to claim 1, characterized in that: The E-NET algorithm sets three detection heads, namely P3 / 8 detection head, P4 / 16 detection head and P5 / 32 detection head; The detection feature map size corresponding to the P3 / 8 detection head is 80*80, which is used to detect targets larger than 8*8. The detection feature map size of the P4 / 16 detection head is 40*40, which is used to detect targets larger than 16*16. The detection feature map size corresponding to the P5 / 32 detection head is 20*20, which is used to detect targets larger than 32*32.

8. The method for removing soil blocks for a potato harvester based on E-NET according to claim 7, characterized in that: The detection head includes a bounding box prediction branch and a classification prediction branch; The bounding box prediction branch includes two convolutional layers, a two-dimensional convolution and a bounding box loss module; The classification prediction branch includes two depth-wise separable convolutions, two convolutional layers, a two-dimensional convolution, and a classification loss module.

Images