YOLO-TI-based tea tender shoot identification method and device

Abstract

The invention relates to a tea tender shoot identification method and device based on YOLO-TI. The method comprises the following steps: S1, constructing a data set; s2, a tea tender shoot recognition model based on YOLO-TI is built; s3, target detection is optimized in combination with an algorithm, and the detection precision is improved; s4, the improved YOLO-TI model is used for identifying and detecting the tender shoots of the tea leaves; and S5, testing and evaluating the YOLO-TI model after the training is completed. The invention relates to the technical field of tea tender shoot recognition. According to the invention, accurate identification of tea tender shoot targets in multiple scenes can be realized, and the probability of false detection and missing detection is reduced. While the detection precision of the model is improved, the generalization of the model in different scenes is enhanced, and the method has important practical significance for intelligent tea picking and development of the tea industry at present.

Description

Technical Field

[0001] The present invention relates to the technical field of tea bud recognition, and in particular to a tea bud recognition method and device based on YOLO-TI. Background Art

[0002] Picking tea buds is a key link in the tea processing process. At present, manual tea picking is labor-intensive, inefficient and costly; working conditions are poor when the weather is bad, and manual picking is slow, which makes it difficult to meet the needs of large-scale tea gardens, resulting in affected tea quality. Existing tea picking machines mostly use a shearing reciprocating cutting method, which lacks selectivity and is prone to missing or picking the wrong tea buds, and may even damage the tea buds. Intelligent recognition and automatic picking technology uses machine vision and robotics technology to achieve efficient and low-cost picking and improve economic benefits. However, the core difficulty of intelligent picking lies in the accurate identification of the position of tea buds. Tea trees are affected by environmental factors, and the height of the buds is inconsistent, which increases the difficulty of identification. Therefore, the development of intelligent recognition technology that can accurately identify the position of tea buds in complex environments is the key to achieving accurate picking. Summary of the invention

[0003] In view of the problems of efficiency and accuracy in tea bud identification in the prior art, the purpose of the present invention is to provide a tea bud identification algorithm based on YOLO-TI, which can realize accurate identification of tea buds through deep learning, thereby improving picking efficiency, saving picking time and reducing labor costs.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] 1. A tea bud recognition method based on YOLO-TI, comprising the following steps: S1. Use image acquisition equipment to obtain tea bud images and construct a training data set; S2. Build a tea bud recognition model based on YOLO-TI, specifically: 1) In the backbone network, use YOLOv10s as the baseline model, use the GELAN module to replace the C2f module in the baseline model, and improve the SPPF module to build the SPPELAN module; introduce the MLLA module to optimize the PSA module in the baseline model and build the PSAMLLA module; use the multi-threading idea to change the backbone network of the model to a multi-threaded structure; 2) In the neck network, build the LSKA attention mechanism module, and the data of the neck network is output to the classification head through the LSKA module; 3) Construct a new bounding box loss function Enhanced_IoU to replace the original bounding box loss function; S3, using the training data set to train the constructed tea bud recognition model to obtain a trained tea bud recognition model; S4. Use the trained tea bud recognition model to identify the tea buds. Furthermore, the specific method of S1 is: Use image acquisition equipment to obtain images of tea buds, and the image data is collected from images of tea buds in tea gardens under different weather conditions, different occlusions, and different distribution conditions; The image data is resized, all images are resized to 1280x720 uniform pixels, image processing technology is used to remove noise in the image, and contrast enhancement processing is performed on the image data to obtain a training data set. Furthermore, the specific implementation method of the LSKA module is: The global context information of the image is captured by horizontal and vertical 1D large kernel depth convolution, and the calculation formula is: In the formula, y represents the output feature map, x represents the input feature map, and W 1x1 , They represent the weight matrices of different convolution kernels respectively, σ represents the activation function, and the σ activation function uses the Rectified Linear Unit, and its formula is: σ(x)=max(0,x). Furthermore, the specific implementation method of the GELAN module and the SPPELAN module is: Using depth-separable convolution operation, the calculation formula is: y=σ(W 1x1 *σ(W 3x3 *x)+σ(W 5x5 *x)) In the formula, y represents the output feature map, x represents the input feature map, and W 1x1 , W 3x3 , W 5x5 They represent the weight matrices of different convolution kernels, σ represents the activation function, * represents the convolution operation, and the formula for depth-separable convolution is: y=σ(W d *σ(W p *x)) Among them, W d , W p Represent the weights of depthwise convolution and pointwise convolution respectively; The pyramid pooling method is used to extract multi-scale feature information and fuse them. Furthermore, the specific implementation method of the PSAMLLA module is: Integrate multi-level feature representation to enhance the model's detection ability for targets of different scales. The calculation formula is: In the formula, y represents the output feature map, x represents the input feature map, and F i Represents the feature extraction functions at different levels, α i Represents the corresponding weight coefficient, and the formula for multi-level feature fusion is: Among them, β i represents the weight of feature fusion, f i Represents feature extraction functions at different levels. In this structure, the backbone network is designed through a multi-threaded parallel architecture, upgrading the traditional single-thread processing to a multi-threaded processing mode. Each thread runs independently and is responsible for extracting and processing features at different levels. The specific multi-thread implementation is as follows: Thread 1: CSP1 module: First, preliminary feature extraction is performed to enhance the capture of low-level features of tea buds, especially in complex backgrounds. 3×3 depthwise separable convolution module: further extracts and refines features, reducing the amount of computation while maintaining efficient feature extraction. GELAN module: performs multi-scale feature extraction, especially targeting the morphological complexity of tea buds, to enhance the capture of detail information. Improved SPPELAN module: performs fusion processing of multi-scale context information and outputs the extracted feature maps to the subsequent feature fusion layer. C2fCIB module: This module further enhances the representation ability of the model through cross-layer feature fusion while maintaining low computational complexity. Thread 2: Downsampling operation: Reduce the spatial dimension of the input feature map to capture a wider range of contextual information. CSP2 module: Enhances the extraction of mid-level semantic features, enabling the model to have stronger feature expression capabilities at the mid-level scale. PELAN module: Focuses on extracting complex tea bud features to ensure high accuracy under changing backgrounds. SCDOWN module: performs dimensionality reduction processing and outputs feature maps to the feature fusion layer to provide high-quality features for subsequent multi-scale fusion. PSAMLLA module: Integrates multi-level feature representation to enhance the model’s ability to detect tea buds of different scales. Thread three: 1×1 convolution: performs channel compression, removes redundant information, and retains effective features. CSP3 module: deeply mines high-level semantic information and enhances the robustness of tea bud detection. Feature fusion module: combines the extracted high-level features with low-level features to form rich contextual expressions. In the feature fusion module, feature maps from different threads are fused by weighted summation. The formula for feature fusion is: Among them, a j represents the feature fusion weight, Represents the feature maps output by different threads. Through multi-threaded parallel architecture design, the backbone network can achieve efficient feature extraction and processing, greatly improving the speed and accuracy of tea bud detection, while ensuring the generalization ability of the model in complex scenarios. Furthermore, the specific implementation of the Enhanced_IoU module includes: Calculate the Enhanced_IoU value of each sample to determine the difficulty of the sample. The calculation formula is: Among them, Enhanced_IoU represents the weighted IoU value, IoU represents the intersection-over-union ratio, ω 1 ,ω 2 ,ω 3 ,ω 4 ,ω 5 They represent the weights of intersection ratio, center distance, aspect ratio similarity, boundary overlap rate and area ratio respectively. d represents the center distance, s represents the aspect ratio similarity, o represents the boundary overlap rate and r represents the area ratio. The difficult samples are weighted to enhance their contribution in the training process. By adjusting the IoU calculation method, the accuracy of the model for difficult-to-detect targets is improved. The formula for weighted processing is: L = -α(1-p t ) γ log(p t ) Among them, L represents the loss function, α and γ represent the adjustment parameters, and p t Represents a prediction probability. The specific implementation process of YOLO-TI model detection is: YOLO-TI directly predicts bounding boxes and class probabilities through regression methods, which includes the following steps: The input image passes through the feature extraction network to generate a feature map F; Use the anchor box mechanism to generate candidate regions A and calculate the regression parameters of each candidate region: Among them, (x, y) is the center coordinate of the real bounding box, (a x ,a y ) is the center coordinate of the anchor box, w and h are the width and height of the real bounding box, a w and a h is the width and height of the anchor box; Using the training data set to train the constructed tea bud recognition model to obtain a trained tea bud recognition model; Use the trained tea bud recognition model to identify tea buds. The tea bud recognition device based on YOLO-TI includes: Image acquisition module, used to obtain tea bud images and construct training data; The tea bud recognition module includes a detection module, which is used to use the improved YOLO-TI model to perform real-time detection of tea buds and classify them according to the detection results; an optimization module, which is used to continuously optimize target detection in combination with the Enhanced_IoU algorithm to improve detection accuracy; a feature fusion module, which is used to enhance the model's recognition ability for tea buds of different sizes using multi-scale feature fusion technology; and a classification and recording module, which is used to classify and record the tea buds according to preset classification standards after the system detects them.

[0049] The present invention has the following beneficial effects:

[0050] The present invention utilizes the improved YOLO-TI model, integrates a variety of advanced feature processing and target detection modules, and provides an efficient and accurate solution for the identification of tea buds. This model is optimized specifically for the problem of small size of tea buds and difficulty in identification in complex backgrounds, and significantly improves the accuracy and efficiency of identification. This model utilizes the idea of ​​multi-threading, and changes the backbone network into a multi-threaded structure. Through the multi-threaded parallel architecture design, the backbone network can achieve efficient feature extraction and processing, greatly improving the speed and accuracy of tea bud detection, while ensuring the generalization ability of the model in complex scenes. At the same time, by integrating LSKA, GELAN, MLLA and Enhanced-IoU modules, the YOLO-TI model of the present invention can effectively fuse multi-scale features and enhance the expression ability of tea bud features. This integration method not only enhances the accuracy of the model when processing small targets of tea buds, but also improves the robustness to background noise, ensuring that efficient recognition performance can be maintained under different environmental conditions. The model of the present invention finely adjusts the role of each module, so that the entire system effectively improves the recognition speed and accuracy of tea buds without adding additional computing burden. Especially in complex environments, such as occlusion and light changes, the model shows excellent performance and can accurately identify and locate tea buds, thereby significantly reducing the number of mis-picking and missed picking. In addition, the model of the present invention also optimizes the computational efficiency of the recognition process, and greatly reduces the demand for computing resources through a carefully designed algorithm and network structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of an embodiment;

[0052] Figure 2 Schematic diagram of the overall network framework of the YOLO-TI model of the present invention;

[0053] Figure 3 It is a schematic diagram of the LSKA structure in the present invention;

[0054] Figure 4 It is a schematic diagram of the structure of C2f and SPPELAN after GELAN is replaced in the present invention;

[0055] Figure 5 It is a schematic diagram of the PSAMLLA structure in the present invention. DETAILED DESCRIPTION

[0055] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0056] Embodiment 1:

[0057] The present invention provides a tea bud recognition algorithm based on YOLO-TI, which specifically includes the following steps:

[0058] S01. Construction of Dataset

[0059] Data collection: Use a high-resolution camera to collect images at different picking times and lighting conditions in the tea garden. The specific steps are as follows:

[0060] An RGB camera with a resolution of 1920x1080 was selected and installed on a height-adjustable bracket to ensure that the captured images covered tea buds at different angles and heights;

[0061] Shoot under different lighting conditions such as sunny and cloudy days to ensure data diversity. The acquisition time is selected in different time periods such as early morning, noon and dusk to capture the impact of lighting changes on the image;

[0062] In each shooting, multiple images were taken continuously to ensure that each image contained clear information of the tea buds. In the end, a total of 5,000 tea bud images were collected;

[0063] Manually annotate the collected images, including the location and category information of the tea buds. The specific steps are as follows:

[0064] Use the LabelImg annotation tool to annotate each image. During the annotation process, a rectangular bounding box is used to accurately mark the location of each tea bud;

[0065] Make sure that each marked bounding box tightly surrounds the tea buds, and avoid the bounding box being too large or too small. During the marking process, if the tea buds overlap or are partially blocked, mark each individual bud as accurately as possible;

[0066] After the marking is completed, all the marked data will be checked and corrected to ensure the accuracy and consistency of the marking;

[0067] Perform data augmentation on labeled images to increase data diversity and improve the generalization ability of the model;

[0068] Dataset division: The labeled image dataset is divided into a training set and a validation set in a ratio of 8:2.

[0069] S02. Build a tea bud recognition model based on YOLO-TI;

[0070] Furthermore, select the experimental platform and configure the operating environment of YOLO-TI. The environment configuration of this application includes: downloading the YOLOv10 source code from GitHub, creating a Conda virtual environment, using Python version 3.8, Pytorch version 2.0.0, and CUDA version 11.8. And install according to the configuration file in the source code.

[0071] The data script for configuring the YOLOv10 model with the established data set ensures that the model can read the specified data set.

[0072] The specific method of building the YOLO-TI tea bud recognition model is as follows:

[0073] Configure the parameters required for network model training, such as the number of iterations, optimizer, initial learning rate, final learning rate, weight decay coefficient, momentum parameter, batch size, and input image size.

[0074] Exemplarily, the number of iterations used in this embodiment is 500, the optimizer is SGD, the initial learning rate is 0.01, the final learning rate is 0.0001, the weight decay coefficient is 0.0005, the momentum parameter is 0.9, the batch processing size is 32, and the input image size is 640×640×3.

[0075] YOLOv10s was selected as the baseline model, and improvements were made to improve the accuracy of the model in target recognition, ensure the detection speed of the model, and reduce the missed detection and false detection of tea buds in multiple scenarios and complex environments.

[0076] In the backbone network, the core design module based on the Mamba network structure integrates the linear attention mechanism to build the MLLA module to optimize the PSA module of the model and construct the PSAMLLA module. It uses position encoding to replace the forget gate to provide the necessary position information of the tea buds.

[0077] In the backbone network, the C2f module of YOLOv10 is improved based on the Generalized ELAN module of YOLOv9, and the SPPF module of YOLOv10 is also improved to build the corresponding SPPELAN module. This ensures that the model obtains more effective features while maintaining the inference speed, and improves the efficiency of model feature extraction while retaining more context information.

[0078] In the neck network, the LSKA attention mechanism module is built based on the large core attention mechanism, which reduces the computational and memory costs while maintaining efficient image processing capabilities.

[0079] The Enhanced-IoU written for the features of tea buds is used instead of the original CIoU as the bounding box follow-up function to improve the convergence speed and accuracy of the prediction box.

[0080] The LSKA module captures context information through large-size convolution kernels and generates sparse feature maps. At the same time, through the decomposition and tandem convolution strategies, it reduces the computational and memory costs while maintaining efficient image processing capabilities. The calculation formula is:

[0081]

[0082] In the formula, y represents the output feature map, x represents the input feature map, and W 1x1 , They represent the weight matrices of different convolution kernels respectively, σ represents the activation function, and the σ activation function uses the Rectified Linear Unit, and its formula is:

[0083] σ(x)=max(0,x);

[0084] The output of the large-size convolution kernel is integrated through 1x1 convolution to improve the nonlinear expression ability of the features, and the convolution results are processed through the activation function to further enhance the feature expression.

[0085] The GELAN module optimizes the C2f and SPPF structures, generates high-density feature maps through deep separable convolution and pyramid pooling, and performs feature correction. This ensures that the model obtains more effective features while maintaining the inference speed, improves the efficiency of model feature extraction, and retains more context information. The calculation formula is:

[0086] y=σ(W 1x1 *σ(W 3x3 *x)+σ(W 5x5 *x))

[0087] In the formula, y represents the output feature map, x represents the input feature map, and W 1x1 , W 3x3 , W 5x5 They represent the weight matrices of different convolution kernels, σ represents the activation function, * represents the convolution operation, and the formula for depth-separable convolution is:

[0088] y=σ(W d *σ(W p *x))

[0089] Among them, W d , W p Represent the weights of depthwise convolution and pointwise convolution respectively;

[0090] The GELAN module also includes the following operations:

[0091] y=W 3x3 *(ReLU(W 1x1 *x))

[0092] Among them, ReLU represents the activation function, W 3x3 , W 1x1 Represent 1x1 and 3x3 convolution kernels respectively. Through this module, the nonlinear expression ability of feature maps can be further improved.

[0093] The MLLA module integrates multi-level feature representation to enhance the model's detection ability for targets of different scales. The constructed PSAMLLA module uses position encoding to replace the forget gate, thereby providing the necessary position information of tea buds. Its calculation formula is:

[0094]

[0095] In the formula, y represents the output feature map, x represents the input feature map, and F i Represents the feature extraction functions at different levels, α i Represents the corresponding weight coefficient. This formula can be used to perform weighted fusion of features at different levels. The formula for multi-level feature fusion is:

[0096]

[0097] Among them, β i represents the weight of feature fusion, f i Represents feature extraction functions at different levels. Through this module, multi-level features can be effectively integrated to improve the detection accuracy of small targets such as tea buds. The fusion ratio of features at different levels is adjusted through weight coefficients to optimize the detection effect; the hierarchical feature representation method is used to improve the model's adaptability to small and large targets.

[0098] S03. Combine the Enhanced_IoU algorithm to optimize target detection and improve detection accuracy. The specific steps include:

[0099] The E_IoU algorithm improves the robustness and accuracy of target detection by optimizing the calculation method of IoU and adding factors such as center distance, aspect ratio similarity, and boundary overlap. Its calculation formula is as follows:

[0100]

[0101] Where A and B represent two bounding boxes, |A∩B| represents the intersection area, and |A∩B| represents the union area.

[0102]

[0103] Where (x 1 ,y 1 ) and (x 2 ,y 2 ) represent the center point coordinates of the two bounding boxes respectively.

[0104]

[0105] where w 1 and h 1 Indicates the width and height of the first bounding box, w 2 and h 2 Indicates the width and height of the second bounding box.

[0106]

[0107] Among them, (x 1 ,y 1 ) and (x 2 ,y 2 ) represents the vertex coordinates of the first bounding box, (x 3 ,y 3 ) and (x 4 ,y 4 ) represents the vertex coordinates of the second bounding box.

[0108]

[0109] Where A and B represent the areas of the two bounding boxes respectively.

[0110] Calculate the IoU value of each sample to determine the difficulty of the sample. The calculation formula is:

[0111]

[0112] Among them, Enhanced_IoU represents the weighted IoU value, IoU represents the intersection-over-union ratio, ω 1 ,ω 2 ,ω 3 ,ω 4 ,ω 5 They represent the weights of intersection ratio, center distance, aspect ratio similarity, boundary overlap rate and area ratio respectively. d represents the center distance, s represents the aspect ratio similarity, o represents the boundary overlap rate and r represents the area ratio.

[0113] Perform E_IoU calculation on the filtered detection results to optimize the target detection accuracy.

[0114] The improved YOLO-TI model is used to identify and detect tea buds.

[0115] The input image passes through the feature extraction network to generate a feature map F;

[0116] Use the anchor box mechanism to generate candidate regions A and calculate the regression parameters of each candidate region:

[0117]

[0118]

[0119] Among them, (x, y) is the center coordinate of the real bounding box, (a x ,a y ) is the center coordinate of the anchor box, w and h are the width and height of the real bounding box, a w and a h is the width and height of the anchor box;

[0120] S4. Train the improved YOLO-TI model;

[0121] Based on the experimental platform, the improved YOLO-TI model is trained on the tea bud dataset to obtain the trained network model.

[0122] S5. Test and evaluate the trained YOLO-TI model.

[0123] The YOLO-TI model after training is tested and evaluated, and the specific method is as follows:

[0124] The trained YOLO-TI model is tested based on the test platform, and the detection results are evaluated in terms of precision, recall, average precision, parameter quantity, and FPS.

[0125] Precision (P) refers to the ratio of samples that are actually positive and predicted by the model to samples that are predicted by the model to be positive. Recall (R) refers to the ratio of samples that are actually positive and predicted by the model to samples that are actually positive. Average precision (AP) refers to the operation of integrating the PR curve, which indicates the balance between the precision and recall of the model for a certain category under different thresholds. Mean average precision (mAP) refers to the average of the average precision of all categories, which can reflect the target detection effect of the model. Parameter quantity (Params) refers to the number of parameters contained in the model. The specific calculation formula is as follows:

[0126]

[0127]

[0128]

[0129]

[0130] TP: The number of actual positive predictions as positive, that is, the number of targets correctly detected by the model;

[0131] FN: The number of actual positive classes predicted as negative classes, that is, the number of targets that the model incorrectly detects;

[0132] FP: The number of actual negative classes predicted as positive classes, that is, the number of backgrounds detected by the model as targets.

[0133] Embodiment 2:

[0134] In this example, the improved YOLO-TI model described in Example 1 is compared with the baseline YOLOv10 model and other commonly used models in the tea bud dataset for recognition and detection effects. The results are shown in Table 1.

[0135] Table 1: Comparison of recognition effects between the improved YOLO-TI model and other models

[0136]

[0137] As shown in Table 1, the YOLO-TI model achieved 94.1% in the P index, which is 7.5, 10.1, 7.7, 9.6, 14.0 and 12.7 percentage points higher than YOLOv10s, YOLOv10m, YOLOv9s, YOLOv8s, Fast-RCNN and SSD respectively, significantly better than the accuracy of existing models.

[0138] The mAP0.5 of the YOLO-TI model is 92.8%, which is 17.0 percentage points higher than the two-stage target detection model Fast-RCNN; it is 14.0, 10.7, 8.1, 7.9 and 9.0 percentage points higher than other single-stage target detection models SSD, YOLOv8s, YOLOv9s, YOLOv10m and YOLOv10s, respectively, showing excellent average precision.

[0139] In the parameter comparison, the YOLO-TI model has 10.35M parameters, which is 92.28%, 60.64%, 7.09%, 15.85%, and 23.33% less than Fast-RCNN, SSD, YOLOv8s, YOLOv9s, and YOLOv10m, respectively. Although the number of parameters is slightly more than some YOLOv10s models, it maintains a higher detection accuracy.

[0140] In terms of frame rate (FPS), the FPS of the YOLO-TI model is 185. Although YOLO-TI is slightly inferior to YOLOv10s in frame rate, this is mainly due to the introduction of more feature fusion modules and complex calculation processes, but the improvement in accuracy makes up for the lack of frame rate and still ensures good real-time performance.

[0141] This example demonstrates that the YOLO-TI model can improve recognition accuracy while achieving high performance and low computational complexity, and has high excellence.

[0142] The implementation principle of the above embodiment is:

[0143] S1. Construction of Dataset

[0144] S2. Build a tea bud recognition model based on YOLO-TI

[0145] S3. Combine Enhanced_IoU algorithm to optimize target detection and improve detection accuracy

[0146] S4. The improved YOLO-TI model is used to identify and detect tea buds.

[0147] S5. Test and evaluate the trained YOLO-TI model.

[0148] For the detection of tea buds in laboratory and natural environments, the present invention improves the YOLOv10 model, including the introduction of LSKA (Large Scale Kernel Attention) module, GELAN (Global Enhanced Local Attention Network) module, MLLA (Multi-Level Layer Attention) module and replacement of the bounding box loss function. The improved YOLO-TI model can realize the accurate recognition of tea buds in multiple scenarios, reduce the probability of false detection and missed detection, improve the detection accuracy of the model, enhance the generalization ability of the model, and has important practical significance for the development of intelligent tea picking. In view of the problem that there are requirements for the picking form of tea buds in the production process of famous and high-quality tea, different picking forms are marked separately in the manual marking link, and the trained model can accurately identify different tea bud targets.

Claims

1. A method for identifying young tea buds based on YOLO-TI, It is characterized in that The following steps are involved: S1. Use image acquisition equipment to obtain tea bud images and construct a training data set; S2. Build a tea bud recognition model based on YOLO-TI, specifically: 1) In the backbone network, use YOLOv10s as the baseline model, use the GELAN module to replace the C2f module in the baseline model, and improve the SPPF module to build the SPPELAN module; introduce the MLLA module to optimize the PSA module in the baseline model and build the PSAMLLA module; use the multi-threading idea to change the backbone network of the model to a multi-threaded structure; 2) In the neck network, build the LSKA attention mechanism module, and the data of the neck network is output to the classification head through the LSKA module; 3) Construct a new bounding box loss function Enhanced_IoU to replace the original bounding box loss function; S3, using the training data set to train the constructed tea bud recognition model to obtain a trained tea bud recognition model; S4. Use the trained tea bud recognition model to identify the tea buds.

2. The tea bud recognition method based on YOLO-TI according to claim 1, It is characterized in that The specific method of S1 is: Use image acquisition equipment to obtain images of tea buds, and the image data is collected from images of tea buds in tea gardens under different weather conditions, different occlusions, and different distribution conditions; The image data is resized, all images are resized to 1280x720 uniform pixels, image processing technology is used to remove noise in the image, and contrast enhancement processing is performed on the image data to obtain a training data set.

3. The tea bud recognition method based on YOLO-TI according to claim 1, It is characterized in that The specific implementation method of the LSKA module is: The global context information of the image is captured by horizontal and vertical 1D large kernel depth convolution, and the calculation formula is: In the formula, y represents the output feature map, x represents the input feature map, and W 1x1 , They represent the weight matrices of different convolution kernels, σ represents the activation function, and the σ activation function uses the Rectified Linear Unit, and the formula is: σ(x)=max(0,x).

4. The tea bud recognition method based on YOLO-TI according to claim 3, It is characterized in that The specific implementation method of the GELAN module and the SPPELAN module is: Using depth-separable convolution operation, the calculation formula is: y=σ(W 1x1 *σ(W 3x3 *x)+σ(W 5x5 *x)) In the formula, y represents the output feature map, x represents the input feature map, and W 1x1 , W 3x3 , W 5x5 They represent the weight matrices of different convolution kernels, σ represents the activation function, * represents the convolution operation, and the formula for depth-separable convolution is: y=σ(W d *σ(W p *x)) Among them, W d , W p Represent the weights of depthwise convolution and pointwise convolution respectively; The pyramid pooling method is used to extract multi-scale feature information and fuse them.

5. The tea bud recognition method based on YOLO-TI according to claim 4, It is characterized in that The specific implementation method of the PSAMLLA module is: Integrate multi-level feature representation to enhance the model's ability to detect objects of different scales. The calculation formula is: In the formula, y represents the output feature map, x represents the input feature map, and F i Represents feature extraction functions at different levels, α i Represents the corresponding weight coefficient, and the formula for multi-level feature fusion is: Among them, β i represents the weight of feature fusion, f i Represents feature extraction functions at different levels. The multi-threaded structure is implemented as follows: Thread 1: CSP1 module: First, preliminary feature extraction is performed to enhance the capture of low-level features of tea buds; 3×3 depthwise separable convolution module: further extracts and refines features, reducing the amount of computation while maintaining efficient feature extraction; GELAN module: performs multi-scale feature extraction to enhance the capture of detail information based on the morphological complexity of tea buds; Improved SPPELAN module: performs fusion processing of multi-scale context information and outputs the extracted feature maps to the subsequent feature fusion layer; C2fCIB module: further enhances the representation ability of the model through cross-layer feature fusion; Thread 2: Downsampling operation: reduce the spatial dimension of the input feature map to capture a wider range of contextual information; CSP2 module: Enhances the extraction of mid-level semantic features, making the model more capable of expressing features at the mid-level. PELAN module: focuses on extracting complex tea bud features, ensuring high accuracy under changing backgrounds; SCDOWN module: performs dimensionality reduction processing and outputs feature maps to the feature fusion layer to provide high-quality features for subsequent multi-scale fusion; PSAMLLA module: integrates multi-level feature representation to enhance the model's ability to detect tea buds of different scales; Thread three: 1×1 convolution: performs channel compression, removes redundant information, and retains effective features; CSP3 module: deeply mines high-level semantic information to enhance the robustness of tea bud detection; Feature fusion module: combines the extracted high-level features with low-level features to form rich contextual expressions; In the feature fusion module, feature maps from different threads are fused by weighted summation. The formula for feature fusion is: Among them, a j represents the feature fusion weight, Represents the feature maps output by different threads.

6. The tea bud recognition method based on YOLO-TI according to claim 5, It is characterized in that The specific implementation of the Enhanced_IoU module includes: Calculate the Enhanced_IoU value of each sample to determine the difficulty of the sample. The calculation formula is: Among them, Enhanced_IoU represents the weighted IoU value, IoU represents the intersection-over-union ratio, ω 1 ,ω 2 ,ω 3 ,ω 4 ,ω 5 They represent the weights of intersection ratio, center distance, aspect ratio similarity, boundary overlap rate and area ratio respectively. d represents the center distance, s represents the aspect ratio similarity, o represents the boundary overlap rate and r represents the area ratio. The difficult samples are weighted to enhance their contribution in the training process. By adjusting the IoU calculation method, the accuracy of the model for difficult-to-detect targets is improved. The formula for weighted processing is: L=-α(1-p t ) γ log(p t ) Among them, L represents the loss function, α and γ represent the adjustment parameters, and p t Represents a prediction probability.

7. A tea bud recognition device based on YOLO-TI, based on the tea bud recognition method based on YOLO-TI according to any one of claims 1 to 6, It is characterized in that include: Image acquisition module, used to obtain tea bud images and construct training data; The tea bud recognition module includes a detection module for real-time detection of tea buds using the improved YOLO-TI model and classification based on the detection results; an optimization module for continuously optimizing target detection in combination with the Enhanced_IoU algorithm to improve detection accuracy; The feature fusion module is used to enhance the model's ability to recognize tea buds of different sizes using multi-scale feature fusion technology; the classification and recording module is used to classify and record tea buds according to preset classification standards after the system detects them.

Images