Man-machine cooperation portable fruit thinning system
Through the human-machine collaborative portable fruit removal system, the industrial camera and semi-supervised domain adaptive object detection algorithm are used to realize intelligent assistance from fruit removal operations, solve the problems of high labor costs, different efficiency and quality in the existing fruit removal technologies, and improve the efficiency and fruit quality of fruit removal operations.
Patent Information
- Application Number
- CN202510211627.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2025-06-17
AI Technical Summary
The existing fruit-reducing technology has high labor costs, varying efficiency and quality, and lacks intelligent identification and precise control, resulting in uneven fruit-reducing or damage to fruits, affecting the growth of fruit trees and fruit quality.
It provides a human-machine collaborative portable fruit removal system, including an image acquisition subsystem, a microcomputer subsystem and a visualization subsystem, and uses an industrial camera to collect images, and uses a semi-supervised domain adaptive object detection algorithm to detect and classify young fruits. It transmits fruit removal suggestions through wireless communication technology and visualizes them on the display.
Real-time image acquisition, deep learning algorithm processing, high-speed data transmission and intuitive display of fruit sparse suggestions are realized, which improves the efficiency and quality of fruit sparse operations, reduces labor costs, and realizes the intelligence of fruit sparse.
Smart Images
Figure CN120164209A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of intelligent fruit thinning, and particularly relates to a human-machine collaborative portable fruit thinning system. Background Art
[0002] Common fruit thinning techniques include manual fruit thinning, chemical fruit thinning, and mechanical fruit thinning. Manual fruit thinning: Traditional manual fruit thinning methods can control the timing according to human requirements and thin fruits according to variety characteristics and fruit setting conditions. Chemical fruit thinning: Using chemical agents such as metribuzin and ethephon for flower and fruit thinning can save labor. Mechanical fruit thinning: Utilizing a robot system for fruit thinning in fruit trees, the fruit thinning operation is achieved through a robotic arm and a thinning component, and a crushing component and a collection component are arranged inside for processing the thinned fruits and branches and leaves.
[0003] However, the above fruit thinning techniques have the following disadvantages: 1. High labor costs and low efficiency. Especially in large-scale commercial orchards, there is a problem of high labor intensity, and large-scale operations cannot be completed in time, affecting the fruit thinning effect. 2. Harm to the environment. Improper use of chemical agents will affect fruit quality and the health of fruit trees, and long-term dependence also poses potential hazards to the environment. 3. Uneven thinning and damage to fruits. Lack of intelligent recognition and precise control easily leads to uneven fruit thinning or damage to fruits, affecting the growth of fruit trees and fruit quality.
[0004] Therefore, there is an urgent need for an intelligent fruit thinning method to assist fruit farmers in fruit thinning work to improve the work efficiency of fruit thinning. Summary of the Invention
[0005] In view of the above deficiencies in the prior art, the present invention provides a human-machine collaborative portable fruit thinning system, which solves the problems of high labor costs, inconsistent efficiency and quality, and insufficient intelligence existing in the prior fruit thinning techniques.
[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: On the one hand, the present invention provides a human-machine collaborative portable fruit thinning system, including: an image acquisition subsystem, a microcomputer subsystem, a visualization subsystem, and a power supply subsystem; The image acquisition subsystem is used for image acquisition using an industrial camera. The microcomputer subsystem is used for obtaining the acquired image using wireless communication technology, and processing the acquired image using an object detection algorithm based on semi-supervised domain adaptation to obtain fruit thinning suggestions, and transmitting the fruit thinning suggestions to the visualization subsystem through wireless communication technology. The visualization subsystem is used for displaying the fruit thinning suggestions on a display, providing professional fruit thinning guidance, and collaborating to complete the fruit thinning operation.
[0007] The beneficial effects of the present invention are as follows: Through real-time image acquisition, deep learning algorithm processing, high-speed data transmission, and intuitive display of fruit thinning suggestions, the present invention provides professional technical support to fruit farmers during fruit thinning operations, improves the efficiency of fruit thinning operations, reduces labor costs, improves efficiency and quality, and realizes the intelligence of fruit thinning; By using a display, the fruit thinning suggestions are visualized, improving the intuitiveness of the man-machine collaborative portable fruit thinning system, enhancing the user experience, enabling fruit farmers to directly receive professional fruit thinning guidance, and improving the convenience and accuracy of fruit thinning operations.
[0008] Further, the image acquisition subsystem specifically is: An industrial camera is used to capture pictures of immature fruits to obtain the acquired images.
[0009] The beneficial effects of the above further solution are: The present invention uses a high-resolution industrial camera to collect images of young fruits in the orchard in real time, improving the real-time performance of image data and the high efficiency of processing.
[0010] Still further, the microcomputer subsystem includes: A vision library module, used to connect the industrial camera and set a timed photo for the industrial camera; An outdoor WIFI module, based on outdoor wireless communication technology equipment, used to provide wireless communication technology, connect the image acquisition subsystem and the visualization subsystem, and transmit the acquired images to the microcomputer subsystem; A detection algorithm module, used to use a target detection algorithm based on semi-supervised domain adaptation to detect and classify young fruits in the acquired images, obtain fruit thinning suggestions according to the young fruit detection and young fruit classification, and transmit the fruit thinning suggestions to the visualization subsystem through wireless communication technology.
[0011] Still further, the specific method steps of the detection algorithm module are as follows: Randomly select from existing open-source data sets, generate an image set on a white background to obtain a first image set, and perform data set annotation on the first image set to obtain the corresponding bounding box set of the first image set and the corresponding class label set of the first image set. According to the images collected by the image acquisition subsystem, preprocess the collected images to obtain a second image set; Combine the first image set, the corresponding bounding box set of the first image set, and the corresponding class label set of the first image set into source domain input data, and use the second image set as target domain input data; Use an offline data augmentation method to augment the target domain input data to obtain augmented target domain input data; Build a student model and a teacher model with the same object detection architecture, input the enhanced target domain input data into the teacher model, and input the source domain input data and the enhanced target domain input data into the student model; Using the teacher model, set a preset threshold, screen out the pseudo-labels in the enhanced target domain input data, combine the data input into the student model and the pseudo-labels, and use the distillation loss to train the student model to obtain a guided student model; According to the guided student model, use the soft-EMA mechanism to update the weights of the teacher model; By using the teacher model to guide the student model and using the student model to update the teacher model, construct an object detection algorithm model based on semi-supervised domain adaptation; Use the object detection algorithm model based on semi-supervised domain adaptation to process the collected images, obtain the young fruit detection results, and get the young fruit classification; Combining the young fruit detection results and the young fruit classification, obtain the fruit thinning suggestions, and transmit the fruit thinning suggestions to the visualization subsystem through wireless communication technology.
[0012] Furthermore, the preprocessing of the collected images is specifically as follows: uniformly adjust the size of the collected images to 640×640 and perform normalization processing.
[0013] Furthermore, the offline data augmentation method is specifically as follows: cut and recombine the images of the target domain input data, and perform data augmentation by increasing the number of input data.
[0014] Furthermore, the step of using the teacher model, setting a preset threshold, screening out the pseudo-labels in the enhanced target domain input data, combining the data input into the student model and the pseudo-labels, and using the distillation loss to train the student model to obtain a guided student model is specifically as follows: Use the teacher model to filter the predicted bounding boxes according to the object confidence by non-maximum suppression, and set the intersection over union threshold and the class score threshold, and screen out the bounding boxes with class scores higher than the threshold as pseudo-labels; According to the bounding box and class information of the target domain image and the pseudo-labels input into the student model, use a filter to obtain the distillation loss function, and use the distillation loss function to train the student model to obtain a guided student model.
[0015] The beneficial effects of the above further solution are as follows: The present invention uses a microcomputer for fast image processing, realizes human-machine collaborative work, and improves the efficiency of image processing; combined with deep learning algorithms, especially the semi-supervised domain adaptation object detection algorithm, an offline data augmentation method is proposed and the EMA mechanism is improved to achieve accurate object detection of the collected images, providing professional fruit thinning suggestions for fruit farmers; using wireless communication technology, fast transmission and synchronization of image data are realized, ensuring that the fruit thinning suggestions can be quickly fed back to fruit farmers and improving the response speed of fruit thinning operations.
[0016] To achieve the above object, according to the second aspect of the present invention, a human-machine collaborative portable fruit thinning device is provided, which is applied to a human-machine collaborative portable fruit thinning system, and is characterized by including: a helmet, an industrial camera, a central control device, an edge computing device, a bracket, and a display; The industrial camera, the central control device, and the edge computing device are fixed on the helmet through a bracket; the central control device is connected to the industrial camera and the display respectively through wireless communication technology; The central control device encapsulates: a Raspberry Pi 4B microcomputer, an uninterruptible power supply, and an outdoor wireless communication technology device.
[0017] The beneficial effects of the above solution are as follows: Through an integrated hardware architecture, the present invention realizes the efficient collaborative work between fruit farmers and fruit thinning equipment, significantly improves the operation efficiency and fruit quality, and reduces the labor cost; using the Raspberry Pi 4B microcomputer and carrying the object detection algorithm based on semi-supervised domain adaptation, the present invention can accurately detect and locate young fruits, improving the uniformity and accuracy of thinning; through modular design, it can flexibly cope with the variability of outdoor environments and meet diverse application requirements, adapting to the standardization levels of different orchards; using a display to intuitively display fruit thinning suggestions improves the user experience and enables fruit farmers to directly receive professional fruit thinning guidance, realizing real-time data processing and feedback. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is the system framework diagram of the present invention.
[0019] Figure 2 It is the system structure diagram in this embodiment.
[0020] Figure 3 It is the work flow chart in this embodiment.
[0021] Figure 4 It is the framework diagram of the object detection algorithm model based on semi-supervised domain adaptation in this embodiment.
[0022] Figure 5 It is the flow chart of the offline data augmentation method in this embodiment.
[0023] Figure 6 This is the relational graph of formula symbols in the object detection algorithm based on semi-supervised domain adaptation in this embodiment.
[0024] Figure 7 This is the design diagram of the human-machine collaborative fruit thinning device in this embodiment.
[0025] Among them, 1. Helmet, 2. Industrial camera, 3. Central control device, 4. Edge computing device, 5. Bracket, 6. Display. Specific implementation manners
[0026] The following describes the specific implementation manners of the present invention to facilitate those skilled in the art of this technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation manners. For those of ordinary skill in the art of this technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.
[0027] Before describing this embodiment, the following terms are first explained: Raspberry Pi 4B: The main board of the Raspberry Pi 4B microcomputer; IPS: In-Plane Switching liquid crystal panel technology; OpenCV: Open Source Computer Vision Library; UPS: Uninterruptible Power Supply; NMS: Non-Maximum Suppression; EMA mechanism: A parameter smoothing technique in deep learning, which stabilizes the model performance by calculating the exponentially decaying average of the parameters; MobileNet-SSD: A deep learning object detection model that combines the lightweight architecture of MobileNet and the single-shot multi-box detection mechanism of SSD; EfficientDet: An object detection model that combines a weighted bidirectional feature pyramid network and a compound scaling method.
[0028] Embodiment 1 As Figure 1 shown, the present invention provides a human-machine collaborative portable fruit thinning system, including: an image acquisition subsystem, a microcomputer subsystem, a visualization subsystem, and a power supply subsystem; The image acquisition subsystem is used to acquire images by using an industrial camera.
[0029] In this embodiment, the image acquisition subsystem, by connecting the industrial camera, is used to capture pictures of immature fruits by using the industrial camera to obtain the collected RGB images.
[0030] The microcomputer subsystem is used to acquire the collected images by using wireless communication technology, and process the collected images by using an object detection algorithm based on semi-supervised domain adaptation to obtain fruit thinning suggestions, and transmit the fruit thinning suggestions to the visualization subsystem through wireless communication technology. The microcomputer subsystem includes: A vision library module, which is used to connect an industrial camera and set a timed photographing for the industrial camera; An outdoor WIFI module, based on outdoor wireless communication technology equipment, is used to provide wireless communication technology, connect the image acquisition subsystem and the visualization subsystem, and transmit the collected images to the microcomputer subsystem; A detection algorithm module, which is used to detect and classify young fruits in the collected images by using an object detection algorithm based on semi-supervised domain adaptation, obtain fruit thinning suggestions according to the young fruit detection and young fruit classification, and transmit the fruit thinning suggestions to the visualization subsystem through wireless communication technology.
[0031] In this embodiment, the microcomputer subsystem is based on a Raspberry Pi 4B microcomputer and includes: a vision library module, an outdoor WIFI module, and a detection algorithm module; The vision library module uses the open-source machine vision library OpenCV to control the industrial camera, sets the timed photographing function, and through the image acquisition subsystem, acquires the fruit thinning branch images in the natural orchard environment, that is, the collected RGB images; An outdoor WIFI module, based on outdoor wireless communication technology equipment, is used to provide wireless communication technology, connect the image acquisition subsystem and the visualization subsystem, and transmit the images collected by the image acquisition subsystem to the detection algorithm module of the microcomputer subsystem; The detection algorithm module, by deploying an object detection algorithm based on semi-supervised domain adaptation, is used to accurately detect objects in an unknown target domain in the case of few-shot training, complete young fruit detection and young fruit classification, and obtain fruit thinning suggestions.
[0032] The visualization subsystem is used to display the fruit thinning suggestions on a display, provide professional fruit thinning guidance, and cooperate to complete the fruit thinning operation.
[0033] In this embodiment, the visualization subsystem is used to display the fruit thinning suggestions on a display, provide professional fruit thinning guidance, and cooperate to complete the fruit thinning operation.
[0034] In this embodiment, as Figure 2As shown in the figure, the present invention is described from three aspects: the hardware layer, the algorithm layer, and the user layer. The hardware layer includes: a power supply for continuously powering the microcomputer; an industrial camera for acquiring images; a microcomputer for carrying the object detection algorithm based on semi-supervised domain adaptation in the algorithm layer; a display for visualization; and a WIFI module for data communication among the industrial camera, the microcomputer, and the display. The algorithm layer is mainly the object detection algorithm based on semi-supervised domain adaptation. By acquiring the RGB images, it performs young fruit detection and young fruit classification to obtain fruit thinning suggestions, and transmits the fruit thinning suggestions to the display through the WIFI module. In the user layer, the fruit thinning suggestions are visualized using the display to achieve human-machine collaboration.
[0035] In this embodiment, as Figure 3 shown, the image acquisition end corresponds to the image acquisition subsystem, the image processing end corresponds to the microcomputer subsystem, and the result display end corresponds to the visualization subsystem, showing the working process of the human-machine collaborative portable fruit thinning system.
[0036] The specific method steps of the detection algorithm module are as follows: Randomly select from the existing open-source datasets and generate an image set on a white background to obtain the first image set, and perform dataset annotation on the first image set to obtain the corresponding bounding box set of the first image set and the corresponding class label set of the first image set. According to the images collected by the image acquisition subsystem, preprocess the collected images to obtain the second image set; the preprocessing specifically is: uniformly adjust the size of the collected images to 640×640 and perform normalization processing. Combine the first image set, the corresponding bounding box set of the first image set, and the corresponding class label set of the first image set into source domain input data, and use the second image set as the target domain input data. Use the offline data augmentation method to augment the target domain input data to obtain the augmented target domain input data; the offline data augmentation method specifically is: cut and recombine the images of the target domain input data to perform data augmentation by increasing the number of input data. Establish a student model and a teacher model with the same object detection architecture, input the augmented target domain input data into the teacher model, and input the source domain input data and the augmented target domain input data into the student model. Using the teacher model, set a preset threshold to screen out the pseudo-labels in the augmented target domain input data. Combine the data input into the student model and the pseudo-labels, and use the distillation loss to train the student model to obtain the guided student model. Specifically: Using the teacher model, the predicted bounding boxes are filtered according to the object confidence by non-maximum suppression, and the intersection over union (IoU) threshold and the class score threshold are set to select the bounding boxes with class scores higher than the threshold as pseudo-labels; According to the bounding box and class information of the target domain image and pseudo-labels input to the student model, a distillation loss function is obtained using a filter, and the student model is trained using the distillation loss function to obtain a guided student model; According to the guided student model, the weights of the teacher model are updated using the soft-EMA mechanism; By using the teacher model to guide the student model and the student model to update the teacher model, an object detection algorithm model based on semi-supervised domain adaptation is constructed; Using the object detection algorithm model based on semi-supervised domain adaptation to process the collected images, the young fruit detection results are obtained, and the young fruit classification is obtained; Combining the young fruit detection results and the young fruit classification, thinning suggestions are obtained, and the thinning suggestions are transmitted to the visualization subsystem through wireless communication technology.
[0037] In this embodiment, as Figure 4 shown, the microcomputer subsystem mainly performs calculations based on semi-supervised and adaptive object detection. It adopts an object detection model based on the knowledge distillation framework, uses the teacher model to learn the features of the target domain, and guides the training process of the student model, enabling the student model to combine the features learned on the source domain and perform good object detection on the target domain. The specific calculation process is as follows: Determine the input data: Randomly select from existing open-source datasets and generate an image set on a white background to obtain the first image set , and label the first image set to obtain the corresponding bounding box set of the first image set , and the corresponding class label set of the first image set ; According to the collected images, preprocess the collected images, uniformly adjust the size of the collected images to 640×640 and perform normalization processing to obtain the second image set ; Combine the first image set , the corresponding bounding box set of the first image set , and the corresponding class label set of the first image set into source domain input data , and use the second image set as the target domain input data ; The expression of the source domain input data is: , and the expression of the target domain input data is ; Offline data augmentation: On the target domain, offline data augmentation is proposed to augment the input data of the target domain to obtain augmented input data for the target domain ; As Figure 5 shown, the specific principle of offline data augmentation is to cut and then recombine the images in the input data set of the target domain, thereby increasing the number of input data and effectively solving the detection difficulty caused by limited learning features due to scarce samples; Semi-supervised domain adaptation learning: A student model and a teacher model with the same object detection architecture are established. The teacher model receives the augmented input data of the target domain at the input end , and the student model receives the input data of the source domain at the input end and the augmented input data of the target domain , where ; The teacher model filters the predicted bounding boxes according to the object confidence by NMS, and sets the IoU threshold and the class score threshold, and takes the bounding boxes with class scores higher than the threshold as pseudo-labels; at the same time, the student model is partially trained on the target domain samples through the distillation loss to obtain a guided student model; the expression of the distillation loss is as follows: ; Where represents the distillation loss, represents the augmented input data of the target domain of the teacher model, represents the augmented input data of the target domain at the input end of the student model, represents the detection loss function, represents the bounding box of the pseudo-label generated by the teacher model, represents the class information of the pseudo-label generated by the teacher model, represents the filter corresponding to the bounding box of the pseudo-label, represents the filter corresponding to the class information of the pseudo-label; On the other hand, the teacher model updates the weights through the improved EMA mechanism of the student model, the soft-EMA mechanism, and establishes a smoothing coefficient through the soft threshold and the relationship between the gradient angle between the teacher model and the student model; using the soft-EMA mechanism, the teacher model can gradually learn the average performance of the student model during the training process, thereby providing more stable and reliable guidance for the student model during the training process, helping the student model learn the features of the target domain and improving the robustness and generalization ability of the model; the calculation expression of the soft-EMA mechanism is as follows: ; ; ; ; Among them, represents the weight parameters of the teacher model, represents the weight parameters of the student model, represents the smoothing coefficient, represents the gradient of the target domain dataset trained on the teacher model, represents the gradient of the target domain dataset trained on the student model, represents the dot product value of the backpropagated gradients of the teacher model and the student model; Such as Figure 6 shown, it can be seen that based on the data relationship in the object detection algorithm of semi-supervised domain adaptation, the target domain input data after data augmentation, the augmented target domain input data and , the augmented target domain input data is input into the teacher model to obtain the gradient of the target domain dataset trained on the teacher model, and the augmented target domain input data is input into the student model to obtain the gradient of the target domain dataset trained on the student model. The source domain input data is input into the student model to obtain the gradient of the source domain dataset trained on the student model.
[0038] In this embodiment, in the object detection algorithm of semi-supervised domain adaptation, the teacher model and the student model can select lightweight object detection algorithms such as the YOLO series, MobileNet-SSD, and EfficientDet.
[0039] Embodiment 2 In this embodiment, as Figure 7 shown, a human-machine collaborative portable fruit thinning device is provided, including: a helmet 1 on the left side of the figure, an industrial camera 2, a central control device 3, an edge computing device 4, a bracket 5, and a display 6 on the right side of the figure; The industrial camera 2, the central control device 3, and the edge computing device 4 are fixed on the helmet 1 through the bracket 5; the central control device 3 is connected to the industrial camera 2 and the display 6 respectively through wireless communication technology; The central control device 3 encapsulates: a Raspberry Pi 4B microcomputer, an uninterruptible power supply, and an outdoor wireless communication technology device; Industrial camera B uses a 12-megapixel color USB high-definition industrial camera, which can capture clear images of unripe fruits in the orchard, providing a high-quality data source for subsequent image analysis; To ensure the continuous and stable operation of the man-machine collaborative portable fruit thinning device in the outdoor environment, an uninterruptible power supply UPS module with an integrated voltage stabilization module is specially designed to ensure the continuity of power supply; Outdoor WIFI device realizes high-speed and stable connection with the external network, facilitating remote monitoring and data synchronization; The display uses a high-definition IPS touch display, which is designed to be independent of other components, intuitively displays fruit thinning suggestions, improves the user experience, makes the fruit thinning operation more convenient and efficient, ensures that fruit farmers can conveniently view fruit thinning suggestions during operation, and this intuitive display method greatly improves the convenience and accuracy of the operation, thus significantly enhancing the efficiency of the fruit thinning operation and the quality of the fruits; The key control devices are encapsulated in the central control device, made of stainless steel, and achieve good electromagnetic isolation and protection.
Claims
1. A human-machine collaborative portable fruit thinning system, characterized in that: include: Image acquisition subsystem, microcomputer subsystem, visualization subsystem, and power supply subsystem; The image acquisition subsystem is used to acquire images using an industrial camera; The microcomputer subsystem is used to obtain the collected images by using wireless communication technology, and process the collected images by using a target detection algorithm based on semi-supervised domain adaptation to obtain fruit thinning suggestions, and transmit the fruit thinning suggestions to the visualization subsystem by using wireless communication technology; The visualization subsystem is used to display the fruit thinning suggestions on the display, provide professional fruit thinning guidance, and coordinate the fruit thinning operation.
2. The human-machine collaborative portable fruit thinning system according to claim 1, characterized in that: The image acquisition subsystem specifically utilizes an industrial camera to capture pictures of unripe fruits to obtain acquired images.
3. The human-machine collaborative portable fruit thinning system according to claim 1, characterized in that: The microcomputer subsystem comprises: The visual library module is used to connect to industrial cameras and set up scheduled photo taking for them; Outdoor WIFI module, based on outdoor wireless communication technology equipment, is used to provide wireless communication technology, connect the image acquisition subsystem and the visualization subsystem, and transmit the acquired images to the microcomputer subsystem; The detection algorithm module is used to use the target detection algorithm based on semi-supervised domain adaptation to detect and classify young fruits in the collected images, make fruit thinning suggestions based on the young fruit detection and classification, and transmit the fruit thinning suggestions to the visualization subsystem through wireless communication technology.
4. The human-machine collaborative portable fruit thinning system according to claim 3, characterized in that: The specific method steps of the detection algorithm module are as follows: An existing open source data set is randomly selected, and an image set is generated on a white background to obtain a first image set, and a data set annotation is performed on the first image set to obtain a bounding box set corresponding to the first image set and a category label set corresponding to the first image set, and the collected images are preprocessed according to the images collected by the image acquisition subsystem to obtain a second image set; The first image set, the bounding box set corresponding to the first image set, and the category label set corresponding to the first image set are combined as source domain input data, and the second image set is used as target domain input data; Using an offline data enhancement method, the target domain input data is enhanced to obtain enhanced target domain input data; Establish a student model and a teacher model with the same target detection architecture, input the enhanced target domain input data into the teacher model, and input the source domain input data and the enhanced target domain input data into the student model; Using the teacher model, setting a preset threshold, filtering out pseudo labels in the enhanced target domain input data, combining the data and pseudo labels input to the student model, and using the distillation loss to train the student model to obtain a guided student model; According to the guided student model, the weights of the teacher model are updated using the soft-EMA mechanism; By using the teacher model to guide the student model and using the student model to update the teacher model, a target detection algorithm model based on semi-supervised domain adaptation is constructed; The collected images are processed using a target detection algorithm model based on semi-supervised domain adaptation to detect young fruits and obtain young fruit classification. Combining the young fruit detection results and young fruit classification, the fruit thinning suggestions are obtained and transmitted to the visualization subsystem through wireless communication technology.
5. The human-machine collaborative portable fruit thinning system according to claim 4, characterized in that: The preprocessing of the collected images specifically includes: uniformly adjusting the size of the collected images to 640×640 and performing normalization processing.
6. The human-machine collaborative portable fruit thinning system according to claim 4, characterized in that: The offline data enhancement method specifically includes: cutting and reorganizing the image of the target domain input data, and performing data enhancement by increasing the amount of input data.
7. The human-machine collaborative portable fruit thinning system according to claim 4, characterized in that: The teacher model is used to set a preset threshold, filter out pseudo labels in the enhanced target domain input data, combine the data and pseudo labels input to the student model, and use the distillation loss to train the student model to obtain the guided student model, specifically: The teacher model is used to filter the predicted bounding boxes according to the object confidence through non-maximum suppression, and the intersection-over-union ratio threshold and the category score threshold are set to filter out the bounding boxes with category scores higher than the threshold as pseudo labels; According to the bounding box and category information of the target domain image and pseudo-label input to the student model, a distillation loss function is obtained by using a filter, and the student model is trained by using the distillation loss function to obtain a guided student model.
8. A human-machine collaborative portable fruit thinning device, applied to a human-machine collaborative portable fruit thinning system, characterized in that: include: Helmet (1), industrial camera (2), central control device (3), edge computing device (4), bracket (5) and display (6); The industrial camera (2), the central control device (3) and the edge computing device (4) are fixed to the helmet (1) via a bracket (5); the central control device (3) is respectively connected to the industrial camera (2) and the display (6) via wireless communication technology; The central control device (3) encapsulates: a Raspberry Pi 4B microcomputer, an uninterruptible power supply, and outdoor wireless communication technology equipment.