Deep learning-based fresh tea leaf picking grade real-time identification method and system

Abstract

The invention discloses a fresh tea leaf picking grade real-time identification method and system based on deep learning, and relates to the technical field of deep learning, and the method comprises the steps: taking an original picking perception set Imgset as unified perception input, building a structural feature set Strset through a deep learning feature extraction model on the basis of completing image preprocessing, and carrying out the real-time identification of the picking grade of fresh tea leaves; the original discrete bud head and leaf information in the fresh tea leaves is further converted into spatial structure description with a clear connection relation and a hierarchical structure; on the basis, a spatial growth sequence set Seqset conforming to the natural growth rule of the fresh tea leaves is formed by analyzing the continuity of the spatial position, the arrangement direction and the adjacent relation of the bud heads and the leaves at all levels, so that the judgment of the picking level does not only depend on the number of the leaves or local appearance characteristics any more; instead, judgment is carried out based on a complete growth structure, so that the stability of a grade identification result can still be kept even if illumination changes or picking postures change.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a method and system for real-time identification of the picking grade of fresh tea leaves based on deep learning. Background Technology

[0002] As the quality grading requirements for tea picking become increasingly stringent, the ability to quickly and accurately determine the grade of fresh tea leaves at the picking site has gradually become one of the key technical issues in tea production. Especially under the background of large-scale planting and refined management, the picking grade not only directly affects the choice of subsequent processing techniques, but also relates to the quality stability and economic value of the final tea product.

[0003] In existing deep learning-based methods for identifying the picking grade of fresh tea leaves, the identification process typically revolves around visible targets such as buds and leaves, inferring the picking grade by detecting the number, area, or morphological characteristics of leaves. However, fresh tea leaves exhibit a distinct hierarchical structure and growth sequence under natural growth conditions. Different picking grades are not only reflected in the difference in the number of leaves but also in the spatial arrangement and growth continuity between buds and leaves at each level. Existing methods often treat buds and leaves as independent identification targets, lacking a unified model of the overall growth structure of fresh tea leaves. This makes it difficult to accurately reflect the true picking grade when leaves overlap, there is interference from adjacent tea plants, or the picking posture changes. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for real-time identification of the picking grade of fresh tea leaves based on deep learning, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a real-time identification method for the picking grade of fresh tea leaves based on deep learning, comprising the following steps: S1. Collect image data of fresh tea leaves in a natural picking scenario, and preprocess the image data to obtain the original picking perception set Imgset for picking grade identification; S2. Input the original picking perception set Imgset into a pre-trained deep learning feature extraction model to perform joint feature extraction on the bud and leaf regions in the fresh tea leaves, and construct a structural feature set Strset representing the relative positional relationship and structural state of the bud and leaf. S3. Based on the structural feature set Strset, analyze the continuity of buds and leaves at all levels in terms of spatial position, arrangement direction and adjacency relationship, and construct a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves according to the growth order of fresh tea leaves; S4. Based on the spatial growth sequence set Seqset, perform picking grade matching and discrimination, and output the picking grade identification result Gradres.

[0006] Preferably, S1 includes S11; S11. In a natural picking scenario, use the existing image acquisition module set on the picking terminal or picking auxiliary equipment to capture images of the picking area where the fresh tea leaves are located. During the acquisition process, by maintaining the image acquisition module in a fixed installation state or a stable handheld state relative to the tea leaf picking target, the bud area and its adjacent multi-level leaf areas are simultaneously within the imaging field of view of the image acquisition module. Based on the continuous imaging capability of the image acquisition module, multiple frames of images are acquired for the same picking target during the picking process, and the acquired image data are numbered and stored according to the acquisition time sequence. The collected image data is subjected to validity screening to remove image frames that have motion blur, occlusion, or do not contain the complete picking target, and an original image dataset Imgraw representing the original visual information of fresh tea leaves under natural picking conditions is generated.

[0007] Preferably, S1 further includes S12; S12. Based on the original image dataset Imgraw, perform image preprocessing operations on each frame of the image sequentially. The image preprocessing operations include correcting the image brightness distribution, adjusting the image contrast, and suppressing image noise, as well as modulating the boundary clarity between the bud region and the leaf region in the image. After completing the image preprocessing operation, the image size and data format are standardized to ensure that the processed image data meets the input requirements of the deep learning feature extraction model. The preprocessed image data is encapsulated in the order of acquisition to generate the original picking perception set Imgset, which represents the basic perception input in the tea leaf picking scenario.

[0008] Preferably, S2 includes S21; S21. Construct a deep learning feature extraction model with convolutional feature learning as its core; The deep learning feature extraction model includes a convolutional feature extraction unit for extracting local texture features and a feature mapping unit for fusing spatial information; During the model training phase, tea leaf image samples that are consistent with the original picking perception set Imgset in terms of image size, data format, and channel structure are used as training sample inputs, so that the training samples are completely consistent with the input data in the actual inference phase in terms of data structure. By repeatedly learning the visual features corresponding to the bud and leaf regions in the training samples, the deep learning feature extraction model gradually establishes a stable response relationship to the morphological features of the buds, the morphological features of the leaves, and the spatial distribution features of the two. When the feature extraction results on the validation samples remain stable during continuous training iterations and can maintain consistent feature responses across different harvesting scenario samples, the model training is complete, and a pre-trained deep learning feature extraction model Modset is obtained.

[0009] Preferably, S2 further includes S22; S22. Input the original picking perception set Imgset into the pre-trained deep learning feature extraction model Modset in the order of image acquisition, and perform forward feature extraction processing on each frame of image; During the feature extraction process, the pre-trained deep learning feature extraction model Modset generates feature responses at different spatial locations in the image, and forms corresponding feature outputs for the bud region and leaf region respectively, thereby obtaining bud features and leaf features. The bud and leaf features obtained from each frame of the image are recorded and organized according to their spatial relationship in the image, forming an initial feature set of buds and leaves, Feaset, which contains bud features, leaf features, and spatial location information.

[0010] Preferably, S2 further includes S23; S23. Based on the spatial location information contained in the initial feature set Feaset of the bud and leaf, the spatial information between the bud features and each leaf feature is judged one by one. During the judgment process, the relative distance, relative direction, and adjacency relationship between the bud and the leaf are judged respectively. When the bud and any leaf simultaneously meet the preset structural connection conditions in the three types of judgments, it is determined that there is a structural connection relationship between the leaf and the bud. Based on the determined structural connection relationship, the structural state between the bud and multi-level leaves is uniformly organized to generate a set of structural features Strset representing the overall structural state of fresh tea leaves; Structural connection relationships include direct connection relationships, indirect connection relationships, and non-connection relationships.

[0011] Preferably, S3 includes S31; S31. Based on the structural feature set Strset, with the bud as the starting reference node, analyze the continuity of each level of leaves that have a structural connection with the bud in terms of spatial position, arrangement direction and adjacency. During the analysis, the connection relationship between the blades was verified according to the spatial distribution order of the blades, and it was determined whether the adjacent blades maintained continuity in terms of spatial distance, directional changes and adjacency status. Combinations of buds and leaves that meet the spatial continuity condition are marked as the same continuous growth path, providing a sorting basis for the subsequent construction of growth sequence.

[0012] Preferably, S3 further includes S32; S32. Based on the continuous structural path set Pathset, the node order in each continuous structural path is sorted: the nodes in the path are sorted according to the spatial progression from the bud to the leaf to determine the growth order between the bud and the leaves at each level. Each sorted continuous structural path is represented as a complete spatial growth sequence, and all spatial growth sequences are organized in a unified manner to generate a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves.

[0013] Preferably, S4 includes S41; S41. Based on the spatial growth sequence set Seqset, analyze the number of bud and leaf levels, arrangement order and structural integrity in each spatial growth sequence. The growth sequence characteristics obtained from the analysis are matched with the pre-established harvesting grade discrimination rules, and the corresponding harvesting grade is determined based on the matching results. The identified picking grades are output to form the final picking grade recognition result, Gradres.

[0014] A real-time tea leaf picking grade recognition system based on deep learning includes a tea leaf image acquisition module, a tea leaf feature extraction module, a feature analysis and construction module, and a picking grade matching module. The tea leaf image acquisition module acquires image data of fresh tea leaves in a natural picking scenario, and preprocesses the image data to obtain the original picking perception set Imgset for picking grade recognition. The tea feature extraction module inputs the original picking perception set Imgset into a pre-trained deep learning feature extraction model to perform joint feature extraction on the bud and leaf regions of fresh tea leaves, and constructs a structural feature set Strset representing the relative positional relationship and structural state of the bud and leaf. The feature analysis construction module analyzes the continuity of buds and leaves at all levels in terms of spatial position, arrangement direction and adjacency based on the structural feature set Strset, and constructs a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves according to the growth order of fresh tea leaves. The picking grade matching module performs picking grade matching and discrimination based on the spatial growth sequence set Seqset, and outputs the picking grade identification result Gradres.

[0015] This invention provides a method and system for real-time identification of the picking grade of fresh tea leaves based on deep learning, which has the following beneficial effects: (1) By using the original picking perception set Imgset as the unified perception input, and after completing the image preprocessing, a structural feature set Strset is constructed through a deep learning feature extraction model. This further transforms the originally discrete bud and leaf information in fresh tea leaves into a spatial structure description with clear connections and hierarchical structure. On this basis, by analyzing the continuity of buds and leaves at all levels in terms of spatial position, arrangement direction and adjacency, a spatial growth sequence set Seqset that conforms to the natural growth law of fresh tea leaves is formed. This makes the picking grade determination no longer rely solely on the number of leaves or local appearance features, but is based on the complete growth structure. Since the picking grade determination is based on the determined bud validity, leaf level number and structural integrity, the grade recognition result can still be kept stable even when the light changes or the picking posture changes, avoiding frequent fluctuations in the grade during continuous recognition.

[0016] (2) Based on the original picking perception set Imgset, a pre-trained deep learning feature extraction model Modset is constructed and introduced, so that the bud and leaf regions in the tea leaf images no longer exist only in the form of pixels or local textures, but are transformed into feature representations with clear spatial meaning and structural orientation. Specifically, by strictly maintaining the consistency between the training samples and the original picking perception set Imgset in terms of image size, data format and channel structure during the model training stage, the pre-trained deep learning feature extraction model Modset can stably output feature responses consistent with those during the training stage in actual use, avoiding structural recognition deviations caused by inconsistent input distribution. On this basis, forward feature extraction processing is performed on each frame of image, and the bud features and leaf features and their spatial position information are uniformly organized into the initial feature set Feaset for buds and leaves, so that subsequent processing can directly perform structural relationship judgment based on known spatial coordinates. Furthermore, by judging the relative distance, relative direction and adjacency relationship one by one, the connection state between buds and leaves is clearly divided into direct connection relationship, indirect connection relationship and non-connection relationship.

[0017] (3) By using the continuous structural path set Pathset and the further formed spatial growth sequence set Seqset, this scheme transforms the originally dispersed spatial relationship between buds and multi-level leaves in fresh tea leaves into an ordered growth sequence that conforms to the natural growth law, so that the growth structure of the picking object can be fully expressed in a continuous and sortable form. On this basis, instead of directly judging the grade based on the appearance features of the image, the picking grade is strictly matched and output according to the validity of the buds, the number of leaf levels and the structural integrity that have been judged in the spatial growth sequence set Seqset. This avoids misjudgment of the grade caused by local occlusion, changes in leaf posture or interference from adjacent tea plant leaves, so that the picking grade judgment process has clear judgment basis, stable output logic and good on-site adaptability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the steps of the real-time identification method for the picking grade of fresh tea leaves based on deep learning in this invention. Figure 2 This is a schematic diagram of the real-time identification system for the picking grade of fresh tea leaves based on deep learning, as described in this invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1

[0020] This invention provides a deep learning-based method for real-time identification of the picking grade of fresh tea leaves. Please refer to [link / reference]. Figure 1 This includes the following steps: S1. Collect image data of fresh tea leaves in a natural picking scenario, and preprocess the image data to obtain the original picking perception set Imgset for picking grade identification; S2. Input the original picking perception set Imgset into a pre-trained deep learning feature extraction model to perform joint feature extraction on the bud and leaf regions in the fresh tea leaves, and construct a structural feature set Strset representing the relative positional relationship and structural state of the bud and leaf. S3. Based on the structural feature set Strset, analyze the continuity of buds and leaves at all levels in terms of spatial position, arrangement direction and adjacency relationship, and construct a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves according to the growth order of fresh tea leaves; S4. Based on the spatial growth sequence set Seqset, perform picking grade matching and discrimination, and output the picking grade identification result Gradres.

[0021] In this embodiment, the deep learning-based real-time tea leaf picking grade identification method constructed through steps S1 to S4 uses the original picking perception set Imgset as a unified perception input. After image preprocessing, a structural feature set Strset is constructed through a deep learning feature extraction model. This further transforms the originally discrete bud and leaf information in the tea leaves into a spatial structure description with clear connections and hierarchical structure. On this basis, by analyzing the continuity of buds and leaves at each level in terms of spatial position, arrangement direction, and adjacency, a spatial growth sequence set Seqset that conforms to the natural growth law of tea leaves is formed. Thus, the picking grade is no longer determined solely by the number of leaves or local appearance features, but is based on the complete growth structure.

[0022] In real-world natural harvesting scenarios, such as when leaves overlap, leaves from adjacent tea plants enter the field of view, or buds are partially obscured, traditional methods can easily misclassify leaves from different growth nodes as the same harvested plant. This solution, however, uses a spatial growth sequence set (Seqset) to constrain structural continuity, effectively eliminating leaves that are spatially discontinuous with the buds. This ensures that the harvesting grade identification results (Gradres) always correspond to the actual growth state of a single fresh tea leaf. Furthermore, since the harvesting grade determination is based on the validity of the identified buds, the number of leaf layers, and structural integrity, the stability of the grade identification results is maintained even when there are changes in light intensity or harvesting posture, avoiding frequent fluctuations in grade during continuous identification. Therefore, this solution can achieve real-time, stable, and natural-growth-compliant identification of fresh tea leaf harvesting grades without altering the data acquisition hardware, providing a reliable basis for on-site harvesting decisions and subsequent processing grading. Example 2

[0023] Specifically: S1 includes S11; S11. In a natural picking scenario, use the existing image acquisition module set on the picking terminal or picking auxiliary equipment to capture images of the picking area where the fresh tea leaves are located. The image acquisition module is a camera device capable of outputting digital image data; During the acquisition process, by maintaining the image acquisition module in a fixed installation state or a stable handheld state relative to the tea leaf picking target, the bud area and its adjacent multi-level leaf areas are simultaneously within the imaging field of view of the image acquisition module. Based on the continuous imaging capability of the image acquisition module, multiple frames of images are acquired for the same picking target during the picking process, and the acquired image data are numbered and stored according to the acquisition time sequence. The collected image data is subjected to validity screening to remove image frames that have motion blur, occlusion, or do not contain the complete picking target, and an original image dataset Imgraw representing the original visual information of fresh tea leaves under natural picking conditions is generated. It should be noted that: Image acquisition module: refers to the digital camera device originally configured on the picking terminal or picking auxiliary equipment, which can output continuous or discrete image data of fresh tea leaves without changing the hardware structure; The original image dataset Imgraw refers to the collection of fresh tea leaf images obtained directly by the image acquisition module and after validity screening. It represents the original visual state information of fresh tea leaves in a natural picking scenario. By using a multi-frame continuous acquisition method, the adjacent images in the original image dataset Imgraw are made to have temporal continuity, providing the necessary temporal basis for subsequent analysis based on spatial growth sequences.

[0024] S1 further includes S12; S12. Based on the original image dataset Imgraw, perform image preprocessing operations on each frame of the image sequentially. The image preprocessing operations include correcting the image brightness distribution, adjusting the image contrast, and suppressing image noise, as well as modulating the boundary clarity between the bud region and the leaf region in the image. After completing the image preprocessing operation, the image size and data format are standardized to ensure that the processed image data meets the input requirements of the deep learning feature extraction model. The preprocessed image data is encapsulated in the order of acquisition to generate the original picking perception set Imgset, which represents the basic perception input in the tea leaf picking scenario. It should be noted that: Image brightness distribution correction refers to adjusting the overall distribution of pixel brightness values ​​in a single frame image so that the bud area and leaf area maintain a relatively stable brightness performance under different lighting conditions. During implementation, the brightness channel of the image is analyzed to obtain the current brightness distribution state of the image, and the overall brightness of the image is adjusted linearly or non-linearly based on the preset brightness target range to make the brightness value distribution more balanced, thereby avoiding local over-darkness or over-brightness caused by strong light, backlight or shadow; through brightness distribution correction, the same picking target can obtain a relatively consistent visual brightness base under different natural lighting conditions, providing a stable input for subsequent bud and leaf feature extraction; Image contrast adjustment refers to enhancing the grayscale or color differences between different areas in an image, making the bud area, leaf area and background area more visually distinct. During implementation, by expanding or compressing the range of pixel intensity variation in the image, the degree of difference between the bud edge, leaf edge contour and adjacent areas is improved, making the boundary between the bud and multi-level leaves clearer in the image; by adjusting the contrast, the visual hierarchy between different structural areas of fresh tea leaves is enhanced, which helps the deep learning model to accurately extract structural features and reduces recognition errors caused by background interference. Image noise suppression processing refers to the process of reducing random interference information in an image caused by the imaging process, environmental factors, or equipment jitter. During implementation, the impact of high-frequency noise on the overall image structure is reduced by smoothing the image or performing neighborhood consistency analysis, while preserving the main edge information of the bud and leaf regions. Noise suppression is used to reduce the interference of irrelevant details in the image on the feature extraction process, thereby improving the reliability of subsequent structural feature recognition. Image size standardization refers to adjusting images that have undergone brightness correction, contrast adjustment, and noise suppression to have a consistent spatial resolution and pixel size. During implementation, images are scaled or cropped proportionally according to the input requirements of the deep learning feature extraction model to ensure that all images are consistent in width and height dimensions. By unifying image size specifications, it is ensured that images in the original picking perception set Imgset can be input into the deep learning model with a fixed structure, avoiding abnormal model input due to inconsistent sizes. Image data format specifications refer to the standardized processing of image storage structure and data representation. During implementation, image data is converted into a data format that can be directly read by deep learning models, and the image channel order, data type, and numerical range are standardized to ensure that all image data are consistent in structure and representation. Through data format standardization, the original picking perception set Imgset has a unified data structure under different picking scenarios and different collection batches, improving the stability and repeatability of the overall recognition process.

[0025] In this embodiment, through the coordinated processing of steps S11 and S12, this solution achieves stable acquisition and standardized processing of tea leaf image data in natural picking scenarios without changing the hardware structure of the picking terminal or picking auxiliary equipment. Specifically, the image acquisition module performs multi-frame continuous imaging of the same picking target, and removes image frames with motion blur, occlusion, or incomplete targets before generating the original image dataset Imgraw, ensuring that subsequent processing is based only on valid image data with controllable quality. On this basis, image preprocessing operations such as brightness distribution correction, contrast adjustment, and noise suppression are performed on the original image dataset Imgraw, and the image size specifications and data format are further unified to form a structurally consistent and input-stable original picking perception set Imgset. In actual tea garden picking, under conditions such as direct sunlight on sunny days, alternating shade under the trees, or slight shaking of the equipment held by the pickers, unprocessed images often exhibit uneven brightness, blurred boundaries, or noise interference, directly affecting subsequent recognition results. This solution, through systematic constraints on the image acquisition and preprocessing processes, ensures that the original picking perception set (Imgset) obtained under different times and environmental conditions remains consistent in brightness, boundary clarity, and data structure. This significantly reduces the impact of changes in the picking environment on the subsequent deep learning feature extraction process, providing a stable and repeatable perceptual input foundation for the entire real-time recognition process of picking levels. Example 3

[0026] Specifically: S2 includes S21; S21. Construct a deep learning feature extraction model with convolutional feature learning as its core; The deep learning feature extraction model includes a convolutional feature extraction unit for extracting local texture features and a feature mapping unit for fusing spatial information; During the model training phase, tea leaf image samples that are consistent with the original picking perception set Imgset in terms of image size, data format, and channel structure are used as training sample inputs, so that the training samples are completely consistent with the input data in the actual inference phase in terms of data structure. By repeatedly learning the visual features corresponding to the bud and leaf regions in the training samples, the deep learning feature extraction model gradually establishes a stable response relationship to the morphological features of the buds, the morphological features of the leaves, and the spatial distribution features of the two. When the feature extraction results on the validation samples remain stable during continuous training iterations and can maintain consistent feature responses in different picking scenario samples, the model training is completed, and a pre-trained deep learning feature extraction model Modset is obtained. It should be noted that: Convolutional feature learning refers to performing multi-layer convolution operations on the input image, enabling the model to extract structural feature information from the low-level to the high-level image layer by layer. The low-level features correspond to the edge contour information of buds and leaves, while the mid-to-high-level features correspond to the bud region, leaf region and their local morphological combination information. During implementation, the convolutional feature learning process aligns the feature responses of tea leaf images of the same category under different picking postures and lighting conditions, enabling the model to learn feature representations that can distinguish the structure of buds and leaves and are insensitive to environmental changes. When the model can extract feature representations with similar distribution characteristics from tea leaf images of the same structural type under different picking scenarios, the convolutional feature learning process is considered to meet the structural feature extraction requirements. Training sample consistency means that the image samples used to train the deep learning feature extraction model are completely consistent with the original picking perception set Imgset generated in step S12 in terms of image size specifications, data format, channel order and pixel value range. During implementation, before entering the model training stage, the training samples are processed according to the same image preprocessing process as the original picking perception set Imgset, so as to ensure that the data input conditions of the training stage and the actual inference stage are consistent. When the training samples can be used directly as equivalent inputs to the original picking perception set Imgset without additional size adjustment, format conversion or channel rearrangement, the consistency of the training samples is considered to meet the requirements. Model convergence refers to the state in which the feature extraction capability of a deep learning feature extraction model reaches a stable level during continuous training iterations. During implementation, the following methods are used to determine whether the model has reached convergence: During training, fixed validation image samples are periodically selected and input into the model; The feature results output by the model on the validation samples are compared in multiple adjacent training phases; When the output feature results maintain a consistent feature distribution trend in multiple consecutive training stages and no longer change significantly with training iterations, the model's feature extraction capability is considered to have stabilized. When the model's feature output results for the same fresh tea leaf structure remain consistent in multiple consecutive validations, and this consistency can be repeated in samples from different picking scenarios, the model is determined to have reached a stable state that can be used for feature extraction. A deep learning model is defined as a pre-trained deep learning feature extraction model (Modset) when it simultaneously meets the following conditions: The model is able to stably extract structural features of the bud and leaf regions from the input raw harvest perception set Imgset; The model maintains consistent feature outputs for tea leaves of the same structural type across samples from different harvesting environments. After continuous training, the model's feature outputs no longer change significantly with continued training; Once the above conditions are met, the model can be used as a fixed feature extraction module in subsequent steps S22 and S23.

[0027] S2 further includes S22; S22. Input the original picking perception set Imgset into the pre-trained deep learning feature extraction model Modset in the order of image acquisition, and perform forward feature extraction processing on each frame of image; During the feature extraction process, the pre-trained deep learning feature extraction model Modset generates feature responses at different spatial locations in the image, and forms corresponding feature outputs for the bud region and leaf region respectively, thereby obtaining bud features and leaf features. The bud and leaf features obtained from each frame of the image are recorded and organized according to their spatial position in the image, forming an initial feature set of buds and leaves, Feaset, which contains bud features, leaf features, and spatial position information. It should be noted that: Forward feature extraction processing refers to passing the input image sequentially through the feature extraction units of each layer of the pre-trained deep learning feature extraction model Modset, so that the visual information in the image is mapped into a high-dimensional feature representation layer by layer; During implementation, forward feature extraction does not involve updating model parameters, but only performs the feature mapping operation that has been trained in the model, thereby ensuring the consistency of feature extraction results at different time points; when the same image is input multiple times, the feature results output by the pre-trained deep learning feature extraction model Modset remain consistent, and the forward feature extraction process is considered to meet the stability requirements. During feature extraction, the features of the bud region and the leaf region are distinguished by the difference in feature response intensity at different spatial locations in the model. The feature response corresponding to the bud region is concentrated in the area where the bud is located in the image, while the feature response corresponding to the leaf region is distributed in the leaf region around the bud. When the model can stably distinguish the bud feature response region and the leaf feature response region in the same image, and the distinction result is consistent in different harvesting images, the bud and leaf feature distinction is considered effective. The process of constructing the initial feature set Feaset for buds and leaves is a process of uniformly organizing the bud features, leaf features and their corresponding spatial location information extracted from each frame of the image. During implementation, the position coordinates of the bud features in the image and the positional relationship of each leaf feature relative to the bud are recorded for each frame of the image, so that the initial feature set Feaset of the bud and the leaf contains both feature information and spatial position information. Through the construction method, the initial feature set Feaset of the bud and the leaf can be directly used for subsequent spatial relationship analysis, avoiding repeated spatial positioning operations in subsequent steps.

[0028] S2 also includes S23; S23. Based on the spatial location information contained in the initial feature set Feaset of the bud and leaf, the spatial information between the bud features and each leaf feature is judged one by one. During the judgment process, the relative distance, relative direction, and adjacency relationship between the bud and the leaf are judged respectively. When the bud and any leaf simultaneously meet the preset structural connection conditions in the three types of judgments, it is determined that there is a structural connection relationship between the leaf and the bud. Based on the determined structural connection relationship, the structural state between the bud and multi-level leaves is uniformly organized to generate a set of structural features Strset representing the overall structural state of fresh tea leaves; Structural connection relationships include direct connection relationships, indirect connection relationships, and non-connection relationships; It should be noted that: Relative distance judgment: refers to judging the spatial distance between the feature positions of the bud and the feature positions of each leaf based on the spatial position information recorded in the initial feature set Feaset of the bud and leaf. During implementation, the location of the bud feature is used as a spatial reference point to calculate the positional distance of each leaf feature relative to the bud feature, and the distance is compared with the preset structural connection distance range for judgment. Explanation of the criteria for implementation: When the distance between the leaf feature and the bud feature is within the range of the structural connection distance, the leaf is considered to meet the relative distance connection condition; When the distance exceeds this range, the leaf is considered not to belong to the structural composition corresponding to the bud. By judging relative distance, leaves that are too far away from the bud are excluded, thus avoiding the mistaken inclusion of adjacent tea plants or background leaves in the same structure. Relative direction judgment: refers to judging the spatial directional relationship between leaf characteristics and bud characteristics in order to determine whether the leaf is located within a reasonable range of the natural growth direction of the bud; During implementation, the main direction of the bud characteristics is used as a reference direction to determine whether the directional distribution of leaf characteristics relative to the bud characteristics is consistent with the natural unfolding direction of fresh tea leaves. Explanation of the criteria for implementation: When the leaf characteristic distribution direction maintains a continuous or gradually changing relationship with the main direction of the bud, the leaf is considered to meet the relative direction connection condition. When the characteristic direction of the leaf deviates significantly from or is in the opposite direction to the main direction of the bud, the leaf is considered not to meet the structural connection direction condition. By judging relative orientation, we can avoid misidentifying leaves with abnormal orientation or inconsistent posture as part of the same growth structure. Adjacency relationship judgment: refers to judging whether a leaf feature is in a continuous adjacent state with a bud feature or a leaf feature that has been determined to be connected in space; During implementation, the spatial distribution of image features is considered to determine whether there is a significant spatial break between leaf features and bud features or adjacent leaf features. Explanation of the criteria for implementation: When leaf features and bud features or adjacent leaf features are in a continuous spatial distribution relationship, the adjacency relationship condition is considered to be satisfied. When there is a clear spatial distance between the leaf feature and the bud or other leaves, the adjacent relationship condition is not considered to be met. By determining the adjacency relationship, we can ensure that the structural connection relationship is continuous in space and avoid incorrect cross-structure connections. Structural connectivity refers to the spatial connection between buds and leaves, and includes at least the following types: Direct connection relationship: refers to a leaf that simultaneously satisfies the judgment of relative distance, relative direction and adjacency relationship, and directly forms a spatial connection with the bud; Indirect connection: refers to the continuous spatial connection between a leaf and a bud through a leaf that has been identified as having a direct connection relationship; Non-connection relationship: refers to a leaf that does not simultaneously meet the above three judgment conditions and is not included in the structural composition corresponding to the bud; Explanation of the criteria for implementation: When a leaf is determined to have a direct or indirect connection, the leaf is included in the structural connection system corresponding to the bud. When a leaf is determined to be non-connected, it is excluded from the current fresh tea leaf structure. The structural feature set Strset refers to the set formed by uniformly organizing bud features, leaf features and the structural connection relationships, which completely represents the structural state between buds and multi-level leaves in fresh tea leaves. During implementation, the structural feature set Strset explicitly records the connection type between each leaf and bud and its hierarchical position in the structure, providing a stable and determinable input basis for the subsequent construction of the spatial growth sequence set Seqset.

[0029] In this embodiment, through the coordinated processing of steps S21 to S23, this scheme constructs and introduces a pre-trained deep learning feature extraction model Modset based on the original picking perception set Imgset. This transforms the bud and leaf regions in the tea leaf image from merely pixels or local textures into feature representations with clear spatial meaning and structural orientation. Specifically, by strictly maintaining the consistency between the training samples and the original picking perception set Imgset in image size, data format, and channel structure during the model training phase, the pre-trained deep learning feature extraction model Modset can stably output feature responses consistent with those during the training phase during actual use, avoiding structural recognition deviations caused by inconsistent input distributions. Furthermore, by performing forward feature extraction processing on each frame of the image and unifying the bud and leaf features and their spatial location information into an initial feature set Feaset for buds and leaves, subsequent processing can directly determine structural relationships based on known spatial coordinates. Furthermore, by judging the relative distance, relative direction, and adjacency relationship one by one, the connection state between the bud and the leaf is clearly divided into direct connection, indirect connection, and non-connection relationship. Based on this, a structural feature set Strset is generated. Thus, in actual natural harvesting scenarios, even if adjacent tea leaves enter the field of view, leaves occlude each other, or leaf posture changes, it is possible to effectively distinguish between "leaves belonging to the same tea leaf growth structure" and "leaves that are only spatially close in the image but have no growth relationship," avoiding the incorrect inclusion of irrelevant leaves in the same structural system. Therefore, step S2 transforms the tea leaves from a "visual target set" into "growth units with clear structural relationships," providing a reliable, determinable, and reusable structural foundation for the subsequent construction of a spatial growth sequence set Seqset based on the structural feature set Strset. Example 4

[0030] Specifically: S3 includes S31; S31. Based on the structural feature set Strset, with the bud as the starting reference node, analyze the continuity of each level of leaves that have a structural connection with the bud in terms of spatial position, arrangement direction and adjacency. During the analysis, the connection relationship between the blades was verified according to the spatial distribution order of the blades, and it was determined whether the adjacent blades maintained continuity in terms of spatial distance, directional changes and adjacency status. Combinations of buds and leaves that meet the spatial continuity condition are marked as the same continuous growth path, providing a sorting basis for the subsequent growth sequence construction; It should be noted that: Spatial continuity refers to the continuous distribution state between buds and leaves, and between adjacent leaves, in accordance with the natural growth pattern during the process of spatial position changes; In practice, the spatial continuity judgment is based on the following three aspects: The spatial distance between adjacent nodes maintains a progressive relationship. The change in the arrangement direction between adjacent nodes maintains a smooth transition; There are no spatial breaks or jumps between adjacent nodes; When all three conditions are met, it is determined that the corresponding nodes have spatial continuity. The continuous structural path set Pathset refers to the set of node connection paths formed in the structural feature set Strset after spatial continuity judgment, where each path represents a continuous growth relationship path between a bud and several leaves; Each path in the continuous structure path set Pathset contains a clear starting node, connection order, and intermediate node relationships, providing a direct basis for determining the growth order of fresh tea leaves.

[0031] S3 further includes S32; S32. Based on the continuous structural path set Pathset, the node order in each continuous structural path is sorted: the nodes in the path are sorted according to the spatial progression from the bud to the leaf to determine the growth order between the bud and the leaves at each level. Each sorted continuous structural path is represented as a complete spatial growth sequence, and all spatial growth sequences are organized in a unified manner to generate a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves. It should be noted that: Growth sequence refers to the hierarchical relationship formed between buds and leaves at all levels during natural growth; During implementation, the spatial positional relationship between each node and the bud in the continuous structural path is compared to determine the order of the nodes in the path, so that the bud is located at the beginning of the sequence and the remaining leaves are arranged in sequence. When the spatial positional change direction of the nodes in the path is consistent and the hierarchical relationship between the nodes remains stable in multiple samples, the node sorting result is considered to be a valid growth order. The spatial growth sequence set Seqset refers to the set of sequences formed by the continuous structural paths in the continuous structural path set Pathset after the growth order is determined, where each sequence completely represents the natural growth arrangement relationship of the bud and multi-level leaves in space. The spatial growth sequence set Seqset clearly records the hierarchical order and connection structure between buds and leaves, enabling the overall growth structure of fresh tea leaves to be represented in sequence form.

[0032] S4 includes S41; S41. Based on the spatial growth sequence set Seqset, analyze the number of bud and leaf levels, arrangement order and structural integrity in each spatial growth sequence. The growth sequence characteristics obtained from the analysis are matched with the pre-established harvesting grade discrimination rules, and the corresponding harvesting grade is determined based on the matching results. The identified picking grades are output to form the final picking grade recognition result, Gradres. It should be noted that: The criteria for determining the picking grade refer to the criteria for classifying the growth structure of tea buds and leaves based on the spatial growth sequences contained in the spatial growth sequence set Seqset. During implementation, the determination of harvest grade is based on at least the following three determined structural elements: Does the spatial growth sequence contain a clearly defined bud initiation node? The number of leaf layers that form a structural connection with the bud; Does the structural integrity of the spatially grown sequence hold? The above three structural elements all originate from the results already determined in the preceding steps, without introducing new perception or feature calculation processes; During implementation, for each spatial growth sequence in the Seqset set, the harvesting grade is determined in the following order: Determining the effectiveness of buds: Determine whether the spatial growth sequence starts with a bud. When there is no clear bud start node, terminate the level determination of the current sequence; Blade level count statistics: Under the premise that the bud is effective, count the number of leaf levels that form direct and indirect connections with the bud; Structural integrity verification: Determine whether there are structural breaks, hierarchical jumps, or missing connections in the spatial growth sequence; When structural integrity is not met, a valid harvesting grade is not output for the spatial growth sequence; Only when all three of the above steps meet the conditions will the picking grade determination stage begin; The harvesting grade is determined based on the correspondence between the number of buds and leaf layers; During implementation, based on the number of leaf levels subsequently connected to the bud in the spatial growth sequence, the harvesting grade is classified into at least the following types: When only one leaf is connected to the bud, it is determined to be the corresponding low-level picking grade; When two levels of leaves are connected in sequence after the bud, it is determined to be the corresponding mid-level picking grade; When a bud is followed by three or more leaves, it is determined to be the corresponding high-level picking grade; The above-mentioned classification of harvesting grades is based on the determined structural result of the number of leaf layers, and does not depend on image area, color or other appearance parameters; When multiple spatial growth sequences exist in the spatial growth sequence set Seqset, during implementation, the spatial growth sequence with the highest structural integrity and the highest number of leaf levels is selected as the effective growth sequence for the current harvesting target. Based on the harvesting grade determination results of the selected valid growth sequence, the corresponding harvesting grade identification results Gradres are output.

[0033] In this embodiment, through the continuous structural path set Pathset constructed in steps S31 and S32 and the further formed spatial growth sequence set Seqset, this scheme transforms the originally dispersed spatial relationship between buds and multi-level leaves in fresh tea leaves into an ordered growth sequence that conforms to the natural growth law, so that the growth structure of the harvested object can be completely expressed in a continuous and sortable form. On this basis, step S41 does not directly judge the grade based on the image appearance features, but strictly follows the bud validity, leaf level quantity and structural integrity already determined in the spatial growth sequence set Seqset to perform regular matching and output of the harvesting grade, thereby avoiding misjudgment of grade caused by local occlusion, leaf posture changes or interference from adjacent tea plant leaves. In real-world harvesting scenarios, such as when multiple levels of leaves from the same tea plant and overlapping leaves from neighboring tea plants appear simultaneously in the harvesting scene, traditional methods tend to merge multiple leaves based solely on quantity. However, this solution constrains spatial continuity through a continuous structural path set (Pathset), including only leaves that are continuous in spatial distance, arrangement direction, and adjacency relationship into the same growth path. Finally, it clarifies the hierarchical order of buds and leaves through a spatial growth sequence set (Seqset), ensuring that the harvesting grade recognition result (Gradres) always corresponds to the actual growth structure of a single fresh tea leaf. Furthermore, when multiple spatial growth sequences exist, it prioritizes the effective growth sequence with the highest structural integrity and number of leaf levels, avoiding unstable grade results due to local noise or short sequence interference. This provides the harvesting grade discrimination process with clear judgment criteria, stable output logic, and good field adaptability. Example 5

[0034] For a real-time tea leaf picking grade identification system based on deep learning, please refer to [link / reference]. Figure 2 Specifically, it includes a tea leaf image acquisition module, a tea leaf feature extraction module, a feature analysis and construction module, and a picking grade matching module; The tea leaf image acquisition module acquires image data of fresh tea leaves in a natural picking scenario, and preprocesses the image data to obtain the original picking perception set Imgset for picking grade recognition. The tea feature extraction module inputs the original picking perception set Imgset into a pre-trained deep learning feature extraction model to perform joint feature extraction on the bud and leaf regions of fresh tea leaves, and constructs a structural feature set Strset representing the relative positional relationship and structural state of the bud and leaf. The feature analysis construction module analyzes the continuity of buds and leaves at all levels in terms of spatial position, arrangement direction and adjacency based on the structural feature set Strset, and constructs a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves according to the growth order of fresh tea leaves. The picking grade matching module performs picking grade matching and discrimination based on the spatial growth sequence set Seqset, and outputs the picking grade identification result Gradres.

[0035] 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 variations 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 technical solutions and their equivalents.

Claims

1. A method for real-time identification of tea leaf picking grade based on deep learning, characterized in that: Includes the following steps: S1. Collect image data of fresh tea leaves in a natural picking scenario, and preprocess the image data to obtain the original picking perception set Imgset for picking grade identification; S2. Input the original picking perception set Imgset into a pre-trained deep learning feature extraction model to perform joint feature extraction on the bud and leaf regions in the fresh tea leaves, and construct a structural feature set Strset representing the relative positional relationship and structural state of the bud and leaf. S3. Based on the structural feature set Strset, analyze the continuity of buds and leaves at all levels in terms of spatial position, arrangement direction and adjacency relationship, and construct a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves according to the growth order of fresh tea leaves; S4. Based on the spatial growth sequence set Seqset, perform picking grade matching and discrimination, and output the picking grade identification result Gradres.

2. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 1, characterized in that: S1 includes S11; S11. In a natural picking scenario, use the existing image acquisition module set on the picking terminal or picking auxiliary equipment to capture images of the picking area where the fresh tea leaves are located. During the acquisition process, by maintaining the image acquisition module in a fixed installation state or a stable handheld state relative to the tea leaf picking target, the bud area and its adjacent multi-level leaf areas are simultaneously within the imaging field of view of the image acquisition module. Based on the continuous imaging capability of the image acquisition module, multiple frames of images are acquired for the same picking target during the picking process, and the acquired image data are numbered and stored according to the acquisition time sequence. The collected image data is subjected to validity screening to remove image frames that have motion blur, occlusion, or do not contain the complete picking target, and an original image dataset Imgraw representing the original visual information of fresh tea leaves under natural picking conditions is generated.

3. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 2, characterized in that: S1 further includes S12; S12. Based on the original image dataset Imgraw, perform image preprocessing operations on each frame of the image sequentially. The image preprocessing operations include correcting the image brightness distribution, adjusting the image contrast, and suppressing image noise, as well as modulating the boundary clarity between the bud region and the leaf region in the image. After completing the image preprocessing operation, the image size and data format are standardized to ensure that the processed image data meets the input requirements of the deep learning feature extraction model. The preprocessed image data is encapsulated in the order of acquisition to generate the original picking perception set Imgset, which represents the basic perception input in the tea leaf picking scenario.

4. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 3, characterized in that: S2 includes S21; S21. Construct a deep learning feature extraction model with convolutional feature learning as its core; The deep learning feature extraction model includes a convolutional feature extraction unit for extracting local texture features and a feature mapping unit for fusing spatial information; During the model training phase, tea leaf image samples that are consistent with the original picking perception set Imgset in terms of image size, data format, and channel structure are used as training sample inputs, so that the training samples are completely consistent with the input data in the actual inference phase in terms of data structure. By repeatedly learning the visual features corresponding to the bud and leaf regions in the training samples, the deep learning feature extraction model gradually establishes a stable response relationship to the morphological features of the buds, the morphological features of the leaves, and the spatial distribution features of the two. When the feature extraction results on the validation samples remain stable during continuous training iterations and can maintain consistent feature responses across different harvesting scenario samples, the model training is complete, and a pre-trained deep learning feature extraction model Modset is obtained.

5. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 4, characterized in that: S2 further includes S22; S22. Input the original picking perception set Imgset into the pre-trained deep learning feature extraction model Modset in the order of image acquisition, and perform forward feature extraction processing on each frame of image; During the feature extraction process, the pre-trained deep learning feature extraction model Modset generates feature responses at different spatial locations in the image, and forms corresponding feature outputs for the bud region and leaf region respectively, thereby obtaining bud features and leaf features. The bud and leaf features obtained from each frame of the image are recorded and organized according to their spatial relationship in the image, forming an initial feature set of buds and leaves, Feaset, which contains bud features, leaf features, and spatial location information.

6. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 5, characterized in that: S2 also includes S23; S23. Based on the spatial location information contained in the initial feature set Feaset of the bud and leaf, the spatial information between the bud features and each leaf feature is judged one by one. During the judgment process, the relative distance, relative direction, and adjacency relationship between the bud and the leaf are judged respectively. When the bud and any leaf simultaneously meet the preset structural connection conditions in the three types of judgments, it is determined that there is a structural connection relationship between the leaf and the bud. Based on the determined structural connection relationship, the structural state between the bud and multi-level leaves is uniformly organized to generate a set of structural features Strset representing the overall structural state of fresh tea leaves; Structural connection relationships include direct connection relationships, indirect connection relationships, and non-connection relationships.

7. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 6, characterized in that: S3 includes S31; S31. Based on the structural feature set Strset, with the bud as the starting reference node, analyze the continuity of each level of leaves that have a structural connection with the bud in terms of spatial position, arrangement direction and adjacency. During the analysis, the connection relationship between the blades was verified according to the spatial distribution order of the blades, and it was determined whether the adjacent blades maintained continuity in terms of spatial distance, directional changes and adjacency status. Combinations of buds and leaves that meet the spatial continuity condition are marked as the same continuous growth path, providing a sorting basis for the subsequent construction of growth sequence.

8. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 7, characterized in that: S3 further includes S32; S32. Based on the continuous structural path set Pathset, the node order in each continuous structural path is sorted: the nodes in the path are sorted according to the spatial progression from the bud to the leaf to determine the growth order between the bud and the leaves at each level. Each sorted continuous structural path is represented as a complete spatial growth sequence, and all spatial growth sequences are organized in a unified manner to generate a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves.

9. The method for real-time identification of tea leaf picking grade based on deep learning according to claim 8, characterized in that: S4 includes S41; S41. Based on the spatial growth sequence set Seqset, analyze the number of bud and leaf levels, arrangement order and structural integrity in each spatial growth sequence. The growth sequence characteristics obtained from the analysis are matched with the pre-established harvesting grade discrimination rules, and the corresponding harvesting grade is determined based on the matching results. The identified picking grades are output to form the final picking grade recognition result, Gradres.

10. A real-time tea leaf picking grade identification system based on deep learning, applied to the real-time tea leaf picking grade identification method based on deep learning as described in any one of claims 1 to 9, characterized in that: It includes a tea leaf image acquisition module, a tea leaf feature extraction module, a feature analysis and construction module, and a picking grade matching module; The tea leaf image acquisition module acquires image data of fresh tea leaves in a natural picking scenario, and preprocesses the image data to obtain the original picking perception set Imgset for picking grade recognition. The tea feature extraction module inputs the original picking perception set Imgset into a pre-trained deep learning feature extraction model to perform joint feature extraction on the bud and leaf regions of fresh tea leaves, and constructs a structural feature set Strset representing the relative positional relationship and structural state of the bud and leaf. The feature analysis construction module analyzes the continuity of buds and leaves at all levels in terms of spatial position, arrangement direction and adjacency based on the structural feature set Strset, and constructs a spatial growth sequence set Seqset representing the overall growth structure of fresh tea leaves according to the growth order of fresh tea leaves. The picking grade matching module performs picking grade matching and discrimination based on the spatial growth sequence set Seqset, and outputs the picking grade identification result Gradres.

Images