Tea leaf picking point positioning method and system with feature enhancement and adaptive regression

Abstract

The present application relates to the field of computer image processing, in particular to a feature enhancement and self-adaptive regression tea leaf picking point positioning method and system. The basic principle of the method is: firstly, the collected tea bud image is preprocessed for clarity; then, a two-stage model architecture of detection first and then positioning is adopted, an improved YOLOv5 network is used in the detection stage to robustly detect multi-scale buds and suppress background interference; then, the target suitable for picking is screened out and cut into a single bud image; in the positioning stage, an improved YOLOv11-Pose network integrated with an adaptive convolution kernel module is used to accurately regress the picking point coordinates; finally, the coordinates are mapped back to the original image and output. The core technical effect of the present application is: through the synergistic optimization of the two-stage process and the targeted improvement of the model components, the problem of inaccurate picking point positioning and poor robustness caused by image degradation, multi-scale targets, complex background and variable bud morphology in the natural environment is effectively solved, providing a high-precision solution for tea leaf automatic picking.

Description

Technical Field

[0001] This invention relates to the field of computer image processing, specifically to a method and system for locating tea picking points using feature enhancement and adaptive regression. Background Technology

[0002] As an important economic crop, the automation of tea production is a key aspect of the development of smart agriculture. Achieving automated and precise harvesting of tea buds is crucial for addressing seasonal labor shortages and improving tea quality and industry efficiency. Currently, machine vision-based harvesting solutions have replaced traditional indiscriminate cutting machinery and become the mainstream research approach. However, the core challenge lies in how to reliably detect buds and accurately locate their harvesting points in complex natural environments.

[0003] Existing visual localization methods mainly follow two technical paths: one is the indirect localization method of "detection / segmentation + post-processing". Although this method is complex and computationally expensive, it has a high accuracy rate, which is beneficial to improving the efficiency of tea picking. The other is the direct regression method based on key point detection. Although it simplifies the process, the key point features are not obvious due to the small size of tea buds, the thinness of stems, and the variety of postures. This results in insufficient modeling ability for situations such as bending and occlusion, and the localization accuracy is difficult to meet the actual picking requirements.

[0004] Existing technologies face three main bottlenecks: First, image degradation caused by natural lighting, fog, etc., reduces the robustness of detection models; second, the large scale differences and high similarity of buds to the background in images increase the difficulty of accurate detection; and third, the picking point localization model lacks an effective adaptation mechanism for non-rigid morphological changes such as stem bending and tilting. Therefore, there is an urgent need for a solution that can systematically address image quality, multi-scale detection, complex backgrounds, and morphological changes, while possessing high localization accuracy and strong robustness. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a solution for locating tea bud picking points based on YOLO series features enhancement and adaptive regression.

[0006] According to a first aspect of the present invention, a method for locating tea picking points using feature enhancement and adaptive regression is proposed, comprising the following steps: S1. Collect images of tea buds in natural scenes, preprocess and label the tea bud images to obtain the first bud dataset, and divide the first bud dataset into the first training set, the first validation set and the first test set; S2. The first sprout dataset is processed to obtain the second sprout dataset; the second sprout dataset includes the second training set, the second validation set, and the second test set, which correspond to the first training set, the first validation set, and the first test set. S3. Construct an initial two-stage tea bud picking location model, including a bud detection model and a picking point location model, wherein: The bud detection model is based on the YOLOv5 architecture and integrates a multi-scale dilated attention module as well as a spatial and channel reconstruction convolution module; The picking point localization model is based on the YOLOv11-Pose architecture, which integrates adaptive convolutional kernel modules in its backbone and neck networks so that the convolutional kernels can adapt to the shape of the target. S4. Use the second training set and the second validation set to train and optimize the initial two-stage tea bud picking location model in order to determine the final two-stage tea bud picking location model. S5. Using the bud detection model and the second test set, obtain the bud target detection box, and dynamically filter and crop the tea bud targets according to the bud target detection box to obtain the third test set; the images in the third test set are single bud images; S6. Using the picking point location model and the third test set, the two-dimensional picking point coordinates are obtained by regression. S7. Map the coordinates of the two-dimensional picking point back to the image of the tea bud to output the positioning result.

[0007] According to some embodiments, in the method of the first aspect of the present invention, step S2 includes: Based on the MPRNet multi-stage progressive recovery architecture, a multi-stage progressive recovery network MRT-Net is constructed. The first sprout dataset is input into the multi-stage progressive recovery network MRT-Net to obtain the second sprout dataset.

[0008] According to some embodiments, in the method of the first aspect of the present invention, the construction process of the multi-stage progressive recovery network (MRT-Net) includes: A dataset of tea image pairs is constructed for network fine-tuning, wherein each image pair consists of an artificially synthesized degraded tea image and its corresponding reference clear tea image; A pre-trained model with a multi-stage progressive restoration architecture MPRNet is used as the base network. The multi-stage progressive restoration architecture MPRNet includes a first restoration stage, a second restoration stage, and a third restoration stage connected in sequence. The first restoration stage is used to initially restore the overall contour of the image, the second restoration stage is used to perform deeper feature fusion to optimize global semantics, and the third restoration stage is used to repair local details in high-resolution space. Using tea leaf images as a dataset, task-oriented fine-tuning of the pre-trained model was performed to determine the multi-stage progressive recovery network MRT-Net.

[0009] According to some embodiments, in the method of the first aspect of the present invention, a task-oriented fine-tuning process is performed on the pre-trained model: a low learning rate strategy is adopted, and the training objective is to jointly optimize a pixel-level loss function and a loss function sensitive to edge structure.

[0010] According to some embodiments, in the method of the first aspect of the present invention, the dynamic filtering process in step S5 includes: Calculate the area and sharpness of each sprout target detection box, and compare them with preset area thresholds and sharpness thresholds respectively. Filter out targets that simultaneously meet the conditions of having an area greater than the area threshold and a sharpness higher than the sharpness threshold.

[0011] According to some embodiments, in the method of the first aspect of the present invention, the sprout detection model specifically includes: integrating a multi-scale dilated attention module in the neck network of the YOLOv5 architecture to enhance multi-scale feature extraction, and integrating a spatial and channel reconstruction convolution module in its head network to reduce feature redundancy.

[0012] According to some embodiments, in the method of the first aspect of the present invention, the multi-scale hollow attention module operates using a multi-scale attention dispersion mechanism, specifically including: after linear transformation of the input features, distributing them to multiple parallel attention heads, wherein at least one attention head uses a first dilation rate and at least another attention head uses a second dilation rate greater than the first dilation rate, so as to simulate receptive fields of different scales respectively; and finally splicing and fusing the outputs of each attention head.

[0013] According to some embodiments, in the method of the first aspect of the present invention, the spatial and channel reconstruction convolutional module integrated in the head network of the bud detection model in step S3 includes a spatial reconstruction unit and a channel reconstruction unit connected in sequence. The spatial reconstruction unit is used to perform separation and reconstruction operations on the input features to suppress feature redundancy in the spatial dimension. The channel reconstruction unit is used to perform segmentation, transformation, and fusion operations on the features processed by the spatial reconstruction unit in order to compress feature redundancy in the channel dimension.

[0014] According to some embodiments, in the method of the first aspect of the present invention, in the picking point localization model in step S3, the adaptive convolutional kernel module implements adaptive sampling through the following steps: Generate the initial sampling coordinates for the convolution kernel; The deformation offset is predicted by analyzing the input features through auxiliary convolutional layers. Add the deformation offset to the initial sampling coordinates to obtain the new sampling coordinates after deformation; The input features are resampled based on the new sampling coordinates.

[0015] According to a second aspect of the present invention, a tea picking point localization system based on feature enhancement and adaptive regression is proposed, comprising: The data acquisition and processing module is used to acquire and process images of tea buds to obtain a first bud dataset, and divide the first bud dataset into a first training set, a first validation set and a first test set. The image sharpening module is used to sharpen the first sprout dataset to obtain the second sprout dataset; the second sprout dataset includes a second training set, a second validation set, and a second test set corresponding to the first training set, the first validation set, and the first test set. The model building module is used to build an initial two-stage tea bud picking and positioning model, including a bud detection model and a picking point positioning model. The model training and optimization module is used to train and optimize the initial two-stage tea bud picking positioning model using the second training set and the second validation set to obtain the final model. The sprout detection and screening module is used to perform target detection using the sprout detection model and the second test set, and to obtain the third test set through dynamic filtering and cropping. The picking point positioning module is used to perform key point regression on the single bud images in the third test set using the picking point positioning model to obtain two-dimensional picking point coordinates. The result mapping and output module is used to map the coordinates of the two-dimensional picking point back to the original image of the tea bud and output the positioning results.

[0016] The technical effect of this invention lies in its targeted solution to a series of key technical problems in locating tea bud picking points in natural environments, specifically achieved through the following innovations: 1. To address the limitation of existing single-stage models in terms of localization accuracy in complex natural scenes, this invention innovatively adopts a two-stage collaborative architecture of "detection first, localization later." This design significantly improves the recall and accuracy of bud detection by decoupling and sequentially optimizing the tasks of bud detection and picking point regression, and provides high-quality target input for the subsequent localization stage. This results in a substantial improvement in the overall accuracy of tea bud picking point localization, effectively enhancing the efficiency and reliability of automated picking operations.

[0017] 2. To address the issue of image blurring and degradation in natural environments leading to decreased feature quality, this invention employs a multi-stage progressive restoration network based on task fine-tuning for sharpening preprocessing, specifically enhancing the details and contour information of the input image. This significantly improves the input quality for subsequent model processing, fundamentally enhancing robustness to degradation factors such as natural lighting and defocusing.

[0018] 3. To address the issues of missed and false detections caused by the variable scale of young shoots and complex backgrounds, this invention employs an improved YOLOv5 detection model that integrates multi-scale dilated attention (MSDA) and spatial channel reconstruction convolution (SCConv) modules. This enhances the model's multi-scale perception capability and its ability to suppress redundant background features, effectively improving the recall and accuracy of young shoot detection in complex tea garden scenarios.

[0019] 4. To address the issue of inaccurate keypoint regression caused by the slender and varied curves of young shoots and stems, this invention integrates an improved YOLOv11-Pose localization model with an adaptive convolutional kernel AKConv module. This allows the convolutional kernel to dynamically adjust the sampling position according to the target shape. This greatly enhances the modeling capability for irregular shapes, thereby achieving accurate regression of the picking point coordinates.

[0020] Furthermore, the dynamic filtering mechanism proposed in this invention, through dual threshold screening of area and clarity, ensures that only complete and clear buds are sent to the positioning stage, further guaranteeing the reliability of the final positioning result. In summary, through the innovative design of the aforementioned collaborative work, this invention systematically solves key challenges such as image quality, multi-scale detection, complex background interference, and morphological adaptation. It significantly outperforms traditional methods in both detection and positioning accuracy, providing a reliable and efficient visual perception core for automated tea harvesting equipment. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment 1000 of the tea picking point localization method based on feature enhancement and adaptive regression of the present invention. Figure 2 for Figure 1 A flowchart illustrating step S2 in Example 1000; Figure 3 for Figure 2 A schematic diagram of the structure of the multi-stage progressive recovery network 201 in step S2; Figure 4 for Figure 1 A schematic diagram of the structure of the bud detection model 301 constructed in step S3 of Example 1000; Figure 5 for Figure 4 A schematic diagram of the structure of the multi-scale hole attention module MSDA in the bud detection model 301; Figure 6 for Figure 4 A schematic diagram of the spatial and channel reconstruction convolutional module SCConv in the bud detection model 301; Figure 7 for Figure 1 A schematic diagram of the structure of the picking point positioning model 302 constructed in step S3 of Example 1000; Figure 8 for Figure 7 A schematic diagram of the adaptive convolutional kernel module AKConv in the picking point positioning model 302; Figure 9 This is a schematic diagram of an embodiment 2000 of the tea picking point positioning system of the present invention, which is based on feature enhancement and adaptive regression. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Figure 1 This is a flowchart illustrating an embodiment 1000 of the tea picking point localization method based on feature enhancement and adaptive regression according to the present invention. Figure 1 As shown, Example 1000 includes steps S1-S7.

[0024] In step S1, images of tea buds in natural scenes are collected, and the images are preprocessed and labeled to obtain the first bud dataset. The first bud dataset is then divided into the first training set, the first validation set, and the first test set.

[0025] In some specific embodiments, step S1 aims to construct a tea bud image dataset for model training and evaluation, specifically including collecting tea bud images in a natural tea garden environment, preprocessing and labeling the images, and dividing the dataset.

[0026] Optionally, the data collection process should cover the diversity of tea buds under different lighting conditions, growth stages, shading situations, and different tea varieties to ensure data representativeness and model generalization ability. Digital cameras or high-resolution industrial cameras should be used for photography, taking care to avoid strong reflections, shadows, and lens smudges during acquisition. After acquisition, the raw images are initially screened, deleting low-quality images caused by focus failure, motion blur, overexposure, or underexposure, and removing duplicate or invalid images. The screened images are then labeled, with the labeled objects being whole buds suitable for harvesting. The labeling includes: bud target bounding box: the smallest bounding rectangle tightly enclosing a single bud; harvesting point coordinates: a two-dimensional coordinate point located on the bud stem, defining the theoretical position for the robotic arm to perform the harvesting operation.

[0027] Optionally, the specific implementation of step S1 includes: (1) Image Acquisition: In natural tea garden settings, a high-resolution digital camera was used for systematic image acquisition. Shooting was conducted during periods of even daylight (e.g., 9:00-11:00 AM or 2:00-4:00 PM), employing a multi-angle, multi-distance strategy to cover the perspective variations encountered during actual operations. Specific parameters were: a shooting pitch angle covering a range of 30° to 60°, and a shooting distance controlled between 30 cm and 50 cm. After acquisition, images with poor focus or duplicate content were manually removed, ultimately yielding a total of 1026 original tea bud images, primarily of the Longjing spring tea variety.

[0028] (2) Image annotation: The original images were annotated in detail using the Labelme annotation tool. Annotation was performed by trained annotators according to standards established by agronomic experts. Labeling target selection: Only complete, clear, and harvestable tender shoots are labeled.

[0029] Bounding box annotation: Draw a rectangular bounding box that tightly surrounds a single bud.

[0030] Picking point marking: Manually mark a point on the tender shoot stem as the picking point and record its pixel coordinates.

[0031] The annotation results are saved in JSON format files. Each file contains bounding box information and picking point coordinates for all annotated buds in the corresponding image.

[0032] (3) Dataset partitioning: The first bud dataset, consisting of all 1026 labeled images and their corresponding annotation files, was randomly divided into a first training set, a first validation set, and a first test set in a ratio of 8:1:1. That is, the first training set contains approximately 821 images, and the first validation set and the first test set each contain approximately 102 images. A stratified random sampling method was used during the partitioning to ensure that images under different lighting conditions and shooting angles are evenly distributed across the three subsets, in order to avoid model bias caused by uneven data distribution.

[0033] In step S2, the first sprout dataset is sharpened to obtain the second sprout dataset. Optionally, in some specific embodiments, step S2 specifically includes: constructing a multi-stage progressive recovery network MRT-Net based on the MPRNet multi-stage progressive recovery architecture; inputting the first sprout dataset into the multi-stage progressive recovery network MRT-Net to obtain the second sprout dataset. Optionally, the second sprout dataset includes a second training set, a second validation set, and a second test set corresponding to the first training set, the first validation set, and the first test set.

[0034] In step S3, an initial two-stage tea bud picking location model is constructed, including a bud detection model and a picking point location model, wherein: The bud detection model is based on the YOLOv5 architecture and integrates a multi-scale dilated attention module as well as a spatial and channel reconstruction convolution module.

[0035] The picking point localization model is based on the YOLOv11-Pose architecture, which integrates adaptive convolutional kernel modules in its backbone and neck network so that the convolutional kernels can adaptively deform according to the shape of the target.

[0036] In step S4, the initial two-stage tea bud picking location model is trained and optimized using the second training set and the second validation set in the second bud dataset to determine the final two-stage tea bud picking location model.

[0037] Optionally, step S4 aims to systematically train and optimize the parameters of the constructed initial two-stage tea bud picking and localization model (including the bud detection model and the picking point localization model) to obtain the model with the best overall evaluation. This process aims to enable the model to learn the complex mapping from the input image to the target output (bounding box and key points) and maximize its generalization ability on unseen data.

[0038] Preferably, step S4 can be implemented through the following specific embodiments: (1) Training preparation and initialization: Dataset used: Training was performed using the pristine second bud dataset.

[0039] Model initialization: Weights from YOLOv5 and YOLOv11-Pose models pre-trained on a large general dataset were used as the initial parameters for the sprout detection model and the picking point localization model, respectively. This transfer learning strategy accelerates convergence and improves model performance.

[0040] The training data for the sprout detection model is the images in the second validation set and the second training set of the second sprout dataset. Data for the localization model: The dataset used to train the picking point localization model consists of single-bud images obtained by automatically cropping images from the second training set and the second validation set, along with their corresponding picking point coordinate labels.

[0041] (2) Training process and parameter settings: Training epochs: Set the total number of training epochs to 300 to ensure the model learns fully.

[0042] Loss function: Detection model: Optimize the standard composite loss of YOLOv5, including bounding box loss CIoU Loss, target confidence loss, and classification loss.

[0043] Localization model: Optimize the keypoint loss function based on heatmap regression or direct coordinate regression, such as mean squared error loss (MSE Loss) or Wing Loss.

[0044] Optimizer and Learning Rate: The AdamW optimizer was used. The initial learning rate was set to 1e-3, and a cosine annealing scheduling strategy was employed to gradually reduce the learning rate during training for smooth convergence.

[0045] Data augmentation: Online data augmentation is applied during training, including random horizontal flipping, random rotation (±15°), random brightness / contrast adjustment, and Mosaic enhancement (for detection models) to improve model robustness.

[0046] (3) Validation, optimization and model selection: Validation and monitoring: After each training cycle, the model performance metrics are calculated using the validation set. For the detection model, its average accuracy mAP@0.5 is monitored at different intersection-over-union (IoU) thresholds; for the localization model, its normalized mean error (NME) for keypoints is monitored.

[0047] Hyperparameter tuning: Based on the validation set performance, perform small-scale grid search or manual adjustment on key hyperparameters.

[0048] Early stopping and saving: Implement an early stopping strategy; if the validation set metrics do not improve within 20 consecutive epochs, terminate training early. Throughout the training process, always save the model weight file that performs best on the validation set.

[0049] Through the above training process, the optimal bud detection model weight file and picking point positioning model weight file are finally determined and saved, which together constitute the final two-stage tea bud picking positioning model for subsequent actual positioning tasks.

[0050] In step S5, using the bud detection model and the second test set, a bud target detection box is obtained. Based on the bud target detection box, the tea bud targets are dynamically filtered and cropped to obtain a third test set; the images in the third test set are single-bud images. Optionally, the dynamic filtering process in step S5 includes: calculating the area and sharpness of each bud target detection box, comparing them with preset area thresholds and sharpness thresholds respectively, and selecting targets that simultaneously satisfy both the area threshold and the sharpness threshold.

[0051] In some specific embodiments, step S5 utilizes a trained bud detection model to identify all potential bud targets from the sharpened image and obtains high-quality single-bud images through an intelligent screening mechanism, preparing for subsequent precise localization. The specific implementation process is as follows: (1) Batch detection: Input images from the second test set are batch-inputted into the trained improved YOLOv5 sprout detection model for forward inference. The model outputs the bounding box information of all predicted sprouts in each image, including their center coordinates, width, height and confidence score.

[0052] (2) Initial screening based on confidence level: First, the detection boxes are initially screened based on a high confidence threshold to filter out obviously unreliable predictions and obtain a set of candidate sprout targets. .

[0053] (3) Dynamic area filtering: Calculate the area of ​​each candidate box area To accommodate differences in the size and density of buds in different images, an area threshold is used. It is not a fixed value, but dynamically adjusted based on the number of candidate boxes k detected within a single image. Specifically, a base area threshold is set. The formula for calculating the dynamic threshold is: ; in, This is a set maximum reference value. This is an adjustment factor. The design tends to filter out larger targets, more likely to be intact foreground buds, in images with dense buds; a more lenient standard is used in sparse images. Targets that meet the criteria are filtered out. The goal.

[0054] (4) Sharpness Filtering: To exclude blurry targets caused by motion blur or occlusion, the sharpness of the image within each target box after area filtering is calculated. Sharpness can be measured by calculating the variance of the Laplacian operator for the image region. A sharpness threshold is set, and targets with sharpness higher than this threshold are retained. This step ensures that the image input to the localization model has rich texture details.

[0055] (5) Single bud image cropping: Based on the bounding box coordinates of the finally selected high-quality bud targets, the corresponding image region is cropped from the original image. Each cropped image is a single bud image containing a single bud, and the collection of all these single bud images constitutes the third test set.

[0056] In step S6, the two-dimensional coordinates of the picking point are regressed using the picking point localization model and the third test set. Optionally, step S6 uses the trained picking point localization model to perform a refined analysis on each single bud image obtained in step S5, directly regressing the two-dimensional coordinates of the picking point. Specifically, step S6 includes: (1) Model inference: Each single bud image in the third test set is input into the trained improved YOLOv11-Pose picking point localization model. The adaptive convolution kernel module integrated in this model has learned to adjust the sampling points according to the morphology of the bud and stem, thereby accurately capturing key point features.

[0057] (2) Coordinate output: The model outputs a two-dimensional coordinate point (x', y') corresponding to each input single bud image. This coordinate is the picking point position defined in the coordinate system of the single bud image itself.

[0058] In step S7, the coordinates of the two-dimensional picking point are mapped back to the tea bud image to output the localization result. Optionally, step S7 transforms the picking point coordinates output by the localization model in the local coordinate system of the single bud image back to the global coordinate system of the original input image to form a complete localization result. Specifically, step S7 includes: (1) Coordinate mapping: For each picking point coordinate (x', y'), combine it with the original cropping information of its source single bud image - that is, the coordinates of the upper left corner of the bounding box corresponding to the single bud image in the original image. And width and height. Map the local coordinates back to the original image coordinates (x, y) using the following formula: ; This calculation accurately restores the localized result to its correct position in the original scene.

[0059] (2) Result Organization and Output: Finally, the system outputs a structured localization result. For each successfully detected and localized bud in the original input image, the result includes: Sprout ID: A unique identifier.

[0060] The bounding box coordinates in the original image.

[0061] The coordinates of the picking point in the original image are (x, y).

[0062] The results can be directly visualized and overlaid on images for manual inspection, or converted into robot-readable instructions to drive robotic arms to perform automated harvesting operations.

[0063] According to such Figure 1 The technical effects achieved by this solution, as shown in the implementation method, can be summarized as follows: Through a series of targeted designs with synergistic optimization, the key bottleneck problem in the visual positioning of automated tea bud harvesting in natural environments is systematically solved, resulting in a significant improvement in positioning accuracy and robustness. Specifically, this is reflected in: 1. Addressing image degradation issues: By employing a multi-stage progressive restoration network based on tea leaf images for preprocessing, image blurring and contrast reduction caused by natural lighting and fog are effectively overcome, providing high-quality input rich in detail and clear morphological features for subsequent processing, thereby enhancing the system's environmental adaptability from the source.

[0064] 2. Addressing Multi-Scale Detection and Background Interference: During the detection phase, the YOLOv5 model is enhanced by introducing the Multi-Scale Dilated Attention (MSDA) module and the Spatial Channel Reconstruction Convolution (SCConv) module. The MSDA module significantly improves the model's ability to perceive sprout targets with significant scale differences through a multi-head, multi-dilation rate attention mechanism; the SCConv module effectively reduces interference from complex backgrounds by suppressing feature redundancy in spatial and channel dimensions, thus jointly achieving high recall and high accuracy in sprout detection.

[0065] 3. Addressing the issue of inaccurate localization caused by the variable morphology of young shoots: During the localization stage, an adaptive convolutional kernel AKConv module is integrated into the backbone and neck of the YOLOv11-Pose network, enabling the convolutional kernel to dynamically adjust the sampling point position based on input features. This mechanism gives the model a strong adaptability to irregular shapes such as bent and tilted shoot stems, thereby achieving accurate regression of key coordinates of the picking point.

[0066] 4. To address the issue of large errors in the location of sprout picking points: A two-stage collaborative architecture of "detection first, then localization" is adopted, supplemented by a dynamic filtering mechanism based on area and sharpness. This architecture decouples and sequentially optimizes the detection and localization tasks, and the filtering mechanism ensures the target quality of the input localization model, thereby improving the accuracy of sprout picking point localization and enhancing the overall system reliability and efficiency.

[0067] In summary, through the aforementioned deep integration improvements, this invention forms a complete, robust, and high-precision vision solution encompassing image preprocessing, target detection, and keypoint localization, providing reliable technical support for the real-time and precise operation of automated tea-picking equipment.

[0068] Figure 2 for Figure 1 A flowchart illustrating step S2 in Example 1000. (See attached diagram.) Figure 2 As shown, step S2 includes steps S21-S22. Figure 3 for Figure 2 A schematic diagram of the structure of the multi-stage progressive recovery network 201 in step S2.

[0069] In some specific embodiments, in step S21, a multi-stage progressive recovery network MRT-Net is constructed based on the MPRNet multi-stage progressive recovery architecture.

[0070] Optionally, the construction process of the multi-stage progressive recovery network MRT-Net specifically includes: First, a dataset of tea image pairs is constructed for network fine-tuning, where each image pair includes an artificially synthesized degraded tea image and its corresponding clear reference tea image. To simulate image degradation in the complex environment of a tea garden, a hybrid degradation model is designed to synthesize training image pairs. Specifically, for a clear reference image... The degraded version is generated through the following process. : ; in, Standard deviation Gaussian blur, simulating inaccurate focus or slight out-of-focus; Linear motion blur with random angles and lengths simulates light breezes or handheld shooting shakiness; For atmospheric scattering model The slight atomization effect, in which the transmittance Atmospheric light value A is taken as the average of the top 1% of image brightness to simulate morning dew or fog.

[0071] Three degradation operations are superimposed in a random order, and each operation is executed independently with a probability p = 0.7. This hybrid degradation model can more realistically reproduce the image quality degradation caused by the combined effect of multiple degradation factors in tea gardens, especially the blurring of thin stem edges and contrast reduction, providing the network with more challenging and representative training data.

[0072] Secondly, in step S21, a pre-trained model with a multi-stage progressive recovery architecture is used as the base network, such as... Figure 3 As shown, the multi-stage progressive recovery architecture includes a first recovery stage, a second recovery stage, and a third recovery stage connected sequentially, wherein: The first stage is used to initially restore the overall contour of the image. Through feature segmentation and U-Net structure, the overall contour and main structure of the image are quickly restored. The second stage is used to perform deeper feature fusion to optimize global semantics, and to perform deeper multi-scale feature fusion to optimize the global semantic consistency and content coherence of the image. The third stage is used to repair local details in high-resolution space. It operates in high-resolution space, focusing on repairing and enhancing local details such as edges and textures.

[0073] Then, in step S21, the pre-trained model is fine-tuned in a task-oriented manner using tea leaf images on the dataset to determine the multi-stage progressive recovery network (MRT-Net). Optionally, the fine-tuning process includes: employing a low learning rate strategy and using the joint optimization of a pixel-level loss function and an edge-structure-sensitive loss function as the training objective.

[0074] In some specific embodiments, the fine-tuning process in step S21 differs from the traditional method of simply minimizing pixel errors; it employs a joint optimization loss function, whose total loss... Defined as: ; The specific design of each component is as follows: Pixel-level fidelity loss Using Charbonnier loss, ,in, This ensures that the restored image closely resembles the original, clear image at the pixel level.

[0075] Edge structure loss Extracting the predicted image using the Sobel operator and real images Edge map and Calculate the L1 loss: This is to enhance the restoration of the outline of the buds and stems.

[0076] Perceived loss : Utilizing the ReLU layer feature maps of a pre-trained VGG-19 network Calculate the L1 loss between feature maps: This ensures that the restored image is aligned with the clear image at the semantic feature level, enhancing the realism of textures.

[0077] Connectivity constraint loss This is specifically designed to address the issue of easily broken tea stems. Firstly, from a clear image... A rough binary mask for the stem is obtained through Canny edge detection and morphological processing. Then, in restoring the image The local contrast (measured by the standard deviation within the local window) is calculated for the corresponding stem region, and the loss is defined as: ; Where std(I,p) represents the standard deviation within a 3×3 window centered at pixel p. This loss directly encourages the network to produce high-contrast, well-connected restoration results in the stem region, effectively penalizing blur or breaks.

[0078] Optionally, in some specific embodiments, the joint optimization loss function is lost. In, weighting coefficient , , , This is used to balance the differences in the numerical magnitude and importance of the various loss terms to the final task. In a preferred embodiment, we experimentally determined a set of effective weight ratios: setting weights 1.0, =0.5, =0.1, =0.2. This ratio ensures that the network can simultaneously consider pixel fidelity, edge sharpness, semantic consistency, and stem connectivity during the optimization process.

[0079] Step S22: Input the first sprout dataset into the multi-stage progressive recovery network MRT-Net to obtain the second sprout dataset.

[0080] In some specific embodiments, in step S22, the task-enhanced multi-stage progressive recovery network constructed in step S21 is fixed with fixed parameters, and all images from the original first sprout dataset are batch-input into the network for forward inference. The network automatically performs degradation removal and detail enhancement on each input image, outputting a corresponding clear image with significantly improved visual quality. All processed images constitute the second sprout dataset, which shows substantial improvements in overall contrast, sprout outline clarity, and stem connectivity, providing a more reliable and discriminative input for subsequent sprout detection and picking point localization models, thus reducing errors caused by image quality at the source.

[0081] Figure 4 for Figure 1 A schematic diagram of the structure of the bud detection model 301 constructed in step S3 of Example 1000. Figure 5 for Figure 4 A schematic diagram of the structure of the Multiscale Hollow Attention Module (MSDA) in the bud detection model 301. Figure 6 for Figure 4 A schematic diagram of the spatial and channel reconstruction convolutional module SCConv in the bud detection model 301.

[0082] Optionally, such as Figure 4 As shown: In the neck network of the YOLOv5 architecture, the sprout detection model 301 replaces the C3 module in the YOLOv5 architecture with the multi-scale dilated attention module MSDA to enhance multi-scale feature extraction; in the head network of the YOLOv5 architecture, the sprout detection model 301 integrates the spatial and channel reconstruction convolution module SCConv to represent the standard convolution module to reduce feature redundancy.

[0083] In some specific embodiments, tea buds, due to natural factors such as varying lighting and shooting distance / angle, exhibit significant scale differences in the image, placing high demands on the model's scale perception capabilities. In some specific embodiments, such as... Figure 5 As shown, MSDA operates using a multi-scale attention distribution mechanism, which specifically includes: after linear transformation of the input features, it is distributed to multiple parallel attention heads, where at least one attention head uses a first dilation rate and at least another attention head uses a second dilation rate greater than the first dilation rate, so as to simulate receptive fields of different scales respectively; finally, the outputs of each attention head are spliced ​​and fused.

[0084] Optionally, such as Figure 5As shown, the specific implementation of MSDA includes: the input features first pass through multiple parallel linear transformation layers; then they are fed into the "Sliding Window Dilated Attention" module, which divides its channels into different attention heads and applies different dilation rates within each head. For example, r1=1, r2=2, r3=3, so that the receptive fields of each head cover 3×3, 5×5, and 7×7 regions, respectively. The integration of MSDA in this invention allows the model to capture semantic information at different scales, effectively avoiding the loss of details that may be caused by a fixed receptive field, and enhancing the model's ability to express features at multiple scales.

[0085] In the task of detecting tea buds, although traditional convolution operations can extract features from images, the features of tea buds are not obvious and they have a high similarity to the background. Therefore, while the model extracts the target features, it also introduces a lot of redundant information. The bud detection model 301 of this invention uses SCConv instead of standard convolution in the head network.

[0086] like Figure 6 As shown, SCConv is a dedicated processing unit designed to suppress feature redundancy. Optionally, SCConv includes a spatial reconstruction unit (SRU) and a channel reconstruction unit (CRU) connected in sequence. The SRU performs separation and reconstruction operations on the input features to suppress feature redundancy in the spatial dimension; the CRU performs segmentation, transformation, and fusion operations on the features processed by the spatial reconstruction unit to compress feature redundancy in the channel dimension. Figure 6 In the illustrated embodiment, the spatial reconstruction unit (SRU) and channel reconstruction unit (CRU) proposed in this invention work in cascade to jointly optimize the convolutional feature map in both spatial and channel dimensions. This enables the model to focus on key features related to tea buds and effectively improves feature discrimination in complex backgrounds.

[0087] Figure 7 for Figure 1 A schematic diagram of the structure of the picking point positioning model 302 constructed in step S3 of Example 1000. Figure 8 for Figure 7 A schematic diagram of the adaptive convolutional kernel module AKConvSCC in the picking point positioning model 302.

[0088] To address the issue of unstable visual characteristics at the picking point caused by the slender and varied shapes of tea buds and stems, the standard square convolution kernel in the baseline model, due to its fixed geometry, is ill-suited to adapt flexibly to the irregular shapes of the buds, such as bending and tilting, easily leading to positioning errors. Therefore, as... Figure 7 As shown, in this embodiment of the invention, the standard convolutional layers in the backbone network and the neck network are replaced with an adaptive convolutional kernel module AKConv.

[0089] In some specific embodiments, such as Figure 8 As shown, the adaptive convolution kernel module AKConv3021 in this invention achieves adaptive sampling through the following steps: generating initial sampling coordinates for the convolution kernel; analyzing input features through an auxiliary convolution layer to predict deformation offset; adding the deformation offset to the initial sampling coordinates to obtain new sampling coordinates after deformation; and resampling the input features based on the new sampling coordinates.

[0090] In some specific embodiments, the adaptive convolutional kernel module AKConv introduces a deformable sampling mechanism, enabling the standard convolutional kernel to dynamically adjust its sampling position based on the local geometry of the input features. For example... Figure 8 As shown, its specific working process includes: (1) Initial sampling coordinate generation Based on a pre-defined algorithm, a set of initial sampling coordinates is generated for the convolution kernel. , where N is the number of custom sampling points. This algorithm is not limited to generating standard rectangular grids, but can initialize diverse geometric arrangements based on prior knowledge (e.g., for linear stem structures), thus providing a flexible and task-relevant starting point for subsequent deformations.

[0091] (2) Deformation offset prediction A copy of the original feature map with dimensions C×H×W is made and fed into a lightweight auxiliary convolutional layer. This auxiliary layer analyzes the local context of each spatial location on the feature map, such as edge gradients and texture orientation, to predict a set of spatial deformation offsets with dimensions 2N×H×W that correspond one-to-one with the initial sampling coordinates. Here, the 2N dimension corresponds to the offset of N sampling points in the x and y directions.

[0092] (3) The predicted offset With initial sampling coordinates By adding elements one by one, we obtain the dynamically adjusted new sampling coordinates. This step is the core of the module's adaptive capability, enabling the sampling points of the convolution kernel to deviate from their initial positions in real time based on the morphological features of each specific target in the input image, thereby conforming to the actual geometric contour of the target.

[0093] (4) Feature resampling Based on the calculated new sampling coordinates Feature values ​​are sampled on the original input feature map using bilinear interpolation. Because... Typically using floating-point coordinates, bilinear interpolation provides sub-pixel-level feature extraction while ensuring gradient differentiability. This step outputs a feature tensor of dimension N×H×W, where each channel corresponds to a feature response map of a deformable sampling point.

[0094] (5) Feature reshaping and integration The resampled N×H×W feature tensor is reshaped to a valid convolutional input shape. Then, a standard 1×1 convolutional layer is used to weight and fuse the features extracted from different sampling points, transforming the channel dimensions to output the final feature map. This fusion process is equivalent to a parameter-adaptive, shape-variable convolution operation.

[0095] (6) Normalization and activation The fused feature maps are sequentially fed into a batch normalization layer and a SiLU activation function layer for processing to stabilize training and introduce non-linearity, and finally output the enhanced features processed by the AKConv module.

[0096] According to such Figure 7 and Figure 8 The implementation shown in this invention employs a YOLOv11-Pose architecture that integrates AKConv in the backbone and neck networks. This introduces morphological adaptation capabilities in the early and middle stages of feature extraction, allowing subsequent network layers to process morphologically corrected features, fundamentally improving the modeling depth for morphological changes. Furthermore, the initial sampling point arrangement considers the linear morphological prior of tea stems. More importantly, the entire module is trained under an end-to-end loss function driven by the keypoint regression task for tea buds. This ensures that the predicted offset is not a general deformation, but a deformation specifically designed to accurately locate the picking point.

[0097] Through such Figure 7 and Figure 8 The adaptive mechanism of AKConv proposed in the illustrated implementation effectively solves the problems of rigid receptive fields and inaccurate feature extraction in traditional fixed-grid convolution when processing slender, curved, and varied tea buds and stems. This enables the model to significantly improve the feature representation quality of the key picking point, ultimately transforming it into a more accurate and robust picking point coordinate regression result, providing reliable visual positioning assurance for automated picking.

[0098] Figure 9 This is a schematic diagram of an embodiment 2000 of the tea picking point localization system based on feature enhancement and adaptive regression of the present invention. Figure 9As shown, Embodiment 2000 includes a data acquisition and processing module 201, an image enhancement module 202, a model building module 203, a model training and optimization module 204, a bud detection and screening module 205, a picking point positioning module 206, and a result mapping and output module 207.

[0099] The data acquisition and processing module 201 is used to acquire and process images of tea buds to obtain a first bud dataset. In some specific embodiments, the data acquisition and processing module 201 serves as the input end of the system, responsible for acquiring raw tea images and performing standardization processing. Specifically, it captures tea garden scenes using sensors such as cameras, and performs image filtering, standardization, and annotation to construct a structured dataset that can be used for model development, providing a data foundation for all subsequent modules.

[0100] Image sharpening module 202 is used to sharpen the first bud dataset to obtain the second bud dataset. Optionally, as a preprocessing enhancement unit, image sharpening module 202 utilizes a multi-stage progressive recovery network (MRT-Net) specifically fine-tuned for tea images to deblur, dehaze, and enhance details of the input original image. This module aims to improve image quality, especially enhancing the visibility of key morphological features such as buds and stems, to improve the robustness of subsequent modules in degraded environments.

[0101] The model building module 203 is used to build an initial two-stage tea bud picking and positioning model, which includes a bud detection model and a picking point positioning model.

[0102] In some specific embodiments, the model building module 203 serves as the core algorithm carrier, integrating and instantiating the dedicated neural network model required for two-stage localization. Specifically, this includes: constructing an improved YOLOv5 detection model that integrates multi-scale dilated attention and spatial channel reconstruction convolution modules, and an improved YOLOv11-Pose localization model that integrates adaptive convolution kernel modules in the trunk and neck.

[0103] The model training and optimization module 204 is used to train and optimize the initial two-stage tea bud picking localization model using the second training set and the second validation set in the second bud dataset, resulting in the final model. Specifically, the model training and optimization module 204 uses the clarified dataset to drive the detection and localization model through an iterative process of forward propagation, loss calculation, backpropagation, and parameter update. This module continuously fine-tunes the model through validation feedback until it achieves stable and excellent performance, thereby producing a deployable final model.

[0104] The sprout detection and screening module 205 is used to perform target detection on the second test set in the second sprout dataset using a sprout detection model, and obtain the third test set through dynamic filtering and cropping. The picking point localization module 206 is used to perform keypoint regression on the single sprout images in the third test set using a picking point localization model to obtain the two-dimensional picking point coordinates.

[0105] Optionally, the positioning process in this embodiment of the invention is divided into two stages: detection and positioning, specifically including: The bud detection and screening module 205, as the first stage of the localization process, calls the trained detection model to infer the sharpened image and identify all potential bud targets. Subsequently, it applies dynamic filtering rules based on area and sharpness to intelligently screen out complete, clear, high-quality targets and crop them into single-bud images, providing high-quality input for accurate localization. The picking point localization module 206, as the second stage of the localization process, calls the trained localization model to analyze each single bud image. The adaptive convolutional kernel inside the model can dynamically adjust the receptive field according to the bending shape of the bud stem, thereby accurately regressing the two-dimensional coordinates of the picking point in the single bud image.

[0106] The result mapping and output module 207 is used to map the coordinates of the two-dimensional picking points back to the original image and output the positioning results. As the output of the system, the result mapping and output module 207 maps the coordinates of the picking points output by the positioning module on the local single-bud image back to the coordinate system of the original global image through geometric transformation. Finally, the system outputs the bounding box of each bud in the original image and its corresponding picking point coordinates, forming a standardized positioning result that can be directly used for robot path planning. In one embodiment of the present invention, to illustrate the effectiveness of the network construction and strategy adopted in the detection phase, under the same hardware and software configuration and dataset, the YOLOv5 network sequentially employs the MPRNet declarative strategy, MSDA structure, and SCConv structure in the tea bud detection phase to compare the effectiveness of each improvement mechanism. The model is evaluated using detection precision, detection recall, and mAP50. The ablation test results are shown in Table 1.

[0107] Table 1. Ablation experimental results of the improved YOLOv5 model: ; As shown in Table 1, the model's performance on tea bud detection was continuously improved after gradually introducing each improved module. Specifically, after introducing only the image sharpening module, the model's precision, recall, and mAP increased to 91.0%, 90.7%, and 92.0%, respectively, representing improvements of 6.9%, 6.9%, and 5.9% compared to the baseline model. This indicates that image preprocessing effectively improved the quality of the input data, especially enhancing the bud recognition ability in blurry and low-contrast scenes, providing a clearer visual foundation for subsequent detection tasks. Based on this, the introduction of the MSDA multi-scale attention module further improved the three metrics to 92.6%, 92.2%, and 94.1%. The MSDA module enhances the feature perception ability of buds at different scales through a multi-scale dilated attention mechanism, exhibiting particularly robust performance in distant, dense, or occluded scenes, significantly improving the model's adaptability and feature discrimination for multi-scale targets. Finally, after simultaneously introducing three modules, MPRNet, MSDA and SCConv, to construct a complete tea bud detection model, the precision, recall and mAP reached 94.6%, 93.6% and 97.1%, respectively, which were 10.5%, 9.8% and 11% higher than the baseline model.

[0108] To further verify the effectiveness of the improved YOLOv5 model, comparative experiments were conducted with the most commonly used YOLO series models in the current object detection field, including YOLOv5, YOLOv7, YOLOv8, YOLOv9, and YOLOv10, as well as classic detection architectures such as Faster R-CNN, RetinaNet, and RT-DETR. Under the same operating environment, the models were evaluated based on precision, recall, and mAP50. To ensure fairness, all data used the same dataset. The test results of different object detection models on the test set images are shown in Table 2.

[0109] Table 2. Performance comparison of different object detection models: ; As shown in Table 2, the precision of our proposed method reaches 94.6%, which is 10.5 percentage points higher than the best-performing baseline model YOLOv5 (84.1%), and significantly higher than other YOLO series models (YOLOv7~YOLOv10 are all between 80% and 82%). This indicates that our proposed method has stronger accuracy in distinguishing between positive and negative samples, and the false positive rate is significantly reduced. In terms of recall, our proposed method achieves a score of 93.6%, which is better than GroundingDINO (86.4%) and Faster R-CNN (84.5%), which have strong target detection capabilities. This shows that the model can still maintain a high detection rate of sprouts in complex natural scenes, effectively reducing the false negative phenomenon. In terms of the comprehensive performance index mAP@0.5, our proposed method significantly outperforms all the comparison models with a score of 97.1%, which is 11 percentage points higher than the second-place YOLOv5 (86.1%). This fully demonstrates the superiority of our method in overall detection performance, especially its superior robustness when dealing with complex situations such as multi-scale, blurred, and dense backgrounds.

[0110] To further verify the effectiveness of the improved YOLOv11-Pose localization model, it was compared with current mainstream keypoint detection and pose estimation models, including DeepPose based on deep regression, the high-resolution network HRNet, and YOLOv5-Pose, YOLOv8-Pose, and YOLOv11-Pose, which support keypoint detection. Under the same operating environment, keypoint accuracy (PCK) was used as the evaluation metric. PCK@0.5 indicates that the Euclidean distance between the predicted keypoint and the ground truth keypoint is less than 0.5 times the stem diameter, indicating that the keypoint is correctly localized. To comprehensively evaluate the model's performance under different accuracy requirements, experiments compared six different thresholds with varying degrees of leniency, from PCK@0.5 to PCK@1.0. The specific results are shown in Table 3.

[0111] Table 3. Performance comparison of different positioning models: ; As can be seen from the six models in Table 3, the first five models are all one-stage localization models, that is, they directly use keypoint regression methods to locate the picking point based on the input image. The method proposed in this invention is a two-stage model, that is, detection followed by localization. Under the most stringent localization accuracy index PCK@0.5, this method achieves 87.75%, which is significantly better than all the comparison models. Compared with the best-performing baseline model YOLOv11-Pose (68.83%), it is improved by 18.92 percentage points, and compared with YOLOv5-Pose (63.46%) and YOLOv8-Pose (61.54%), it is improved by 24.29 and 26.21 percentage points, respectively. This reflects the fundamental challenge faced by the one-stage method in natural scenes: the model must directly complete the regression of tiny picking points in a single forward propagation, and the network is difficult to achieve accurate localization in complex backgrounds, multi-scale targets, and morphological changes. This experiment further validates the necessity and superiority of the two-stage design in this task, demonstrating stronger robustness in high-precision localization tasks, especially suitable for tea picking scenarios with extremely strict requirements on the location of picking points. As the threshold is gradually relaxed, this method maintains a leading position across all metrics from PCK@0.6 to PCK@1.0. For example, it achieves 91.10% on PCK@0.6, approximately 19.79 percentage points higher than YOLOv11-Pose (71.31%); and 95.92% on PCK@0.8, approximately 14.49 percentage points higher than YOLOv11-Pose (81.43%). This indicates that AKConv, by adaptively adjusting the convolution kernel shape, effectively enhances its ability to extract features from the varied morphologies of tea buds (such as curved stems), thereby significantly improving the accuracy of keypoint localization and the robustness of the model.

[0112] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for locating tea picking points using feature enhancement and adaptive regression, characterized in that, Includes the following steps: S1. Collect images of tea buds in natural scenes, preprocess and label the tea bud images to obtain a first bud dataset, and divide the first bud dataset into a first training set, a first validation set and a first test set; S2. The first sprout dataset is processed to obtain a second sprout dataset; the second sprout dataset includes a second training set, a second verification set, and a second test set corresponding to the first training set, the first verification set, and the first test set. S3. Construct an initial two-stage tea bud picking location model, including a bud detection model and a picking point location model, wherein: The sprout detection model is based on the YOLOv5 architecture and integrates a multi-scale dilated attention module and a spatial and channel reconstruction convolution module. The picking point localization model is based on the YOLOv11-Pose architecture, which integrates adaptive convolutional kernel modules in its backbone and neck networks so that the convolutional kernels can adaptively deform according to the shape of the target. S4. The initial two-stage tea bud picking location model is trained and optimized using the second training set and the second validation set to determine the final two-stage tea bud picking location model. S5. Using the bud detection model and the second test set, obtain the bud target detection box, and dynamically filter and crop the tea bud targets according to the bud target detection box to obtain the third test set; the images in the third test set are single bud images; S6. Using the picking point positioning model and the third test set, the two-dimensional picking point coordinates are obtained by regression. S7. Map the coordinates of the two-dimensional picking point back to the image of the tea bud to output the positioning result.

2. The method according to claim 1, characterized in that, Step S2 includes: Based on the MPRNet multi-stage progressive recovery architecture, a multi-stage progressive recovery network MRT-Net is constructed. The first sprout dataset is input into the multi-stage progressive recovery network MRT-Net to obtain the second sprout dataset.

3. The method according to claim 2, characterized in that, The construction process of the multi-stage progressive recovery network MRT-Net includes: A dataset of tea image pairs is constructed for network fine-tuning, wherein each image pair consists of an artificially synthesized degraded tea image and its corresponding reference clear tea image; A pre-trained model with a multi-stage progressive restoration architecture MPRNet is used as the base network. The multi-stage progressive restoration architecture MPRNet includes a first restoration stage, a second restoration stage and a third restoration stage connected in sequence. The first restoration stage is used to initially restore the overall contour of the image, the second restoration stage is used to perform deeper feature fusion to optimize global semantics, and the third restoration stage is used to repair local details in high-resolution space. Using the tea leaf image dataset, the pre-trained model is fine-tuned in a task-oriented manner to determine the Multi-Stage Progressive Recovery Network (MRT-Net).

4. The method according to claim 3, characterized in that, The task-oriented fine-tuning process for the pre-trained model involves employing a low learning rate strategy and targeting the joint optimization of a pixel-level loss function and a loss function sensitive to edge structures.

5. The method according to claim 1, characterized in that, The dynamic filtering process in step S5 includes: The area and sharpness of each of the sprout target detection boxes are calculated and compared with preset area thresholds and sharpness thresholds, respectively, to filter out targets that simultaneously satisfy the conditions of having an area greater than the area threshold and a sharpness higher than the sharpness threshold.

6. The method according to claim 1, characterized in that, The sprout detection model specifically includes: integrating a multi-scale dilated attention module into the neck network of the YOLOv5 architecture to enhance multi-scale feature extraction, and integrating spatial and channel reconstruction convolution modules into its head network to reduce feature redundancy.

7. The method according to claim 6, characterized in that, The multi-scale hollow attention module operates using a multi-scale attention dispersion mechanism, specifically including: after linear transformation of the input features, it is distributed to multiple parallel attention heads, where at least one attention head uses a first dilation rate and at least another attention head uses a second dilation rate greater than the first dilation rate, so as to simulate receptive fields of different scales respectively; finally, the outputs of each attention head are spliced ​​and fused.

8. The method according to claim 6, characterized in that, The spatial and channel reconstruction convolutional module integrated in the head network of the bud detection model in step S3 includes a spatial reconstruction unit and a channel reconstruction unit connected in sequence. The spatial reconstruction unit is used to perform separation and reconstruction operations on the input features in order to suppress feature redundancy in the spatial dimension. The channel reconstruction unit is used to perform segmentation, transformation and fusion operations on the features processed by the spatial reconstruction unit in order to compress feature redundancy in the channel dimension.

9. The method according to claim 1, characterized in that, In the picking point localization model in step S3, the adaptive convolutional kernel module achieves adaptive sampling through the following steps: Generate the initial sampling coordinates for the convolution kernel; The deformation offset is predicted by analyzing the input features through auxiliary convolutional layers. Add the deformation offset to the initial sampling coordinates to obtain the new sampling coordinates after deformation; The input features are resampled based on the new sampling coordinates.

10. A tea picking point localization system using feature enhancement and adaptive regression, characterized in that, include: The data acquisition and processing module is used to acquire and process images of tea buds to obtain a first bud dataset, and divide the first bud dataset into a first training set, a first validation set and a first test set. An image sharpening module is used to sharpen the first sprout dataset to obtain a second sprout dataset; the second sprout dataset includes a second training set, a second verification set, and a second test set corresponding to the first training set, the first verification set, and the first test set. The model building module is used to build an initial two-stage tea bud picking and positioning model, including the bud detection model and the picking point positioning model. The model training and optimization module is used to train and optimize the initial two-stage tea bud picking positioning model using the second training set and the second validation set to obtain the final model. The sprout detection and screening module is used to perform target detection using the sprout detection model and the second test set, and to obtain a third test set through dynamic filtering and cropping. The picking point positioning module is used to perform key point regression on the single bud images in the third test set using the picking point positioning model to obtain two-dimensional picking point coordinates. The result mapping and output module is used to map the coordinates of the two-dimensional picking point back to the original image of the tea bud and output the positioning result.

Images