A system and method for setting plant extraction parameters based on image processing

Abstract

This invention relates to the field of image processing, and more particularly to a system and method for setting plant extraction parameters based on image processing. The method includes the following steps: deploying multimodal data acquisition sensors on plant raw materials and performing multi-scale data fusion to construct a plant multi-source feature dataset; constructing a three-dimensional feature space matrix of plants based on the plant multi-source feature dataset; and establishing a model for the plant three-dimensional feature space matrix to obtain a plant process parameter prediction model. Therefore, this invention, by fusing multimodal data and reinforcement learning mechanisms, solves the problems of inaccurate parameter prediction, missing component distribution information, and lagging process optimization feedback in traditional plant extraction processes, thereby improving the efficiency and quality control level of plant active ingredient extraction.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a system and method for setting plant extraction parameters based on image processing. Background Technology

[0002] Existing image-processing-based methods for setting plant extraction parameters generally suffer from key technical bottlenecks, including limited utilization of image information, coarse-grained prediction of process parameters, and a disconnect between structure recognition and component correlation. At the image data processing level, most methods rely solely on two-dimensional visible light images or low-resolution RGB images for surface texture analysis, neglecting the deeper information expression of plant tissues in dimensions such as hyperspectral, microstructure, and three-dimensional spatial morphology, resulting in insufficient basic information support for parameter setting. Furthermore, traditional methods primarily rely on expert experience or simple rule mapping in setting parameters such as extraction temperature, time, and solvent type, lacking a modeling logic between effective biological features extracted from images (such as cell density, vascular bundle distribution, and spatial concentration gradient of active ingredients) and the extraction process, leading to a lack of specificity and adaptability in parameter generation. In addition, the image recognition module and subsequent process calculation stages often operate separately, failing to achieve fusion modeling and feedback optimization between image features, extraction sites, component prediction, and thermal response data, making it difficult to address the parameter adjustment needs arising from the differences in tissue structure of different plant raw materials. This hierarchical processing mode also limits the level of intelligence in converting image data into process parameters, affecting the potential for improving extraction efficiency and component yield. Summary of the Invention

[0003] Therefore, it is necessary to provide a plant extraction parameter setting system and method based on image processing to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for setting plant extraction parameters based on image processing is provided, the method comprising the following steps:

[0005] Step S1: Deploy multimodal data acquisition sensors on plant materials and perform multi-scale data fusion to construct a multi-source feature dataset of plants;

[0006] Step S2: Construct a three-dimensional feature space matrix of plants based on a multi-source feature dataset; establish a model for the three-dimensional feature space matrix of plants to obtain a prediction model for plant process parameters;

[0007] Step S3: Acquire real-time plant images; input the real-time plant images into the plant process parameter prediction model to predict plant components, and obtain a spatial distribution prediction map of plant components; perform small-batch extraction experiments to verify the spatial distribution prediction map of plant components, and obtain plant prediction yield data; perform reinforcement learning on the plant prediction yield data and map the model weights to obtain parameter prediction accuracy compensation; iterate the parameter prediction accuracy compensation and the plant prediction yield data to obtain optimized process parameter data.

[0008] Step S4: Modularly package the optimized process parameter data and construct an active ingredient enrichment heat map; overlay the active ingredient enrichment heat map with the optimal extraction sites and provide process parameter suggestions to obtain a plant extraction process decision report.

[0009] The beneficial effects of this invention lie in its ability to comprehensively acquire and express high-dimensional characteristics of plant raw materials through multimodal data acquisition and multi-scale fusion, thereby constructing a three-dimensional feature space matrix covering spectral, morphological, and component spatial distribution. This provides a rich and structured data foundation for the accurate prediction of plant processing parameters. During data processing, real-time plant images are inferred through deep learning using a process parameter prediction model to generate a component spatial distribution prediction map. Combined with prediction yield data obtained from small-batch experimental verification, the model weights are dynamically adjusted using a reinforcement learning algorithm, achieving continuous optimization and adaptive compensation of parameter prediction accuracy. The optimized process parameter data is further modularly packaged into standardized inputs. A heatmap of active ingredient enrichment is constructed using spatial overlay technology, accurately revealing the spatial distribution characteristics of target components in different tissues. Combined with a process parameter prompting system, this effectively guides the selection of extraction sites and the formulation of process plans, ultimately generating a plant extraction process decision report. This report, based on multi-source fusion data and iterative optimization results, provides scientific and quantitative decision support for process execution, reducing the blindness and experience-based dependence of traditional process parameter setting and improving the accuracy and controllability of the extraction process. Therefore, by integrating multimodal data and reinforcement learning mechanisms, this invention solves the problems of inaccurate parameter prediction, missing component distribution information, and lagging feedback in traditional plant extraction processes, thereby improving the efficiency and quality control level of plant active ingredient extraction.

[0010] Preferably, step S1 includes the following steps:

[0011] Step S11: Use a hyperspectral imager with a wavelength range of 400-2500nm and a resolution of 5nm to scan the plant raw materials and collect hyperspectral image data of the raw materials.

[0012] Step S12: Acquire images of plant tissue cell structures using a microscopic imaging system with a magnification of 40-1000X and a pixel accuracy of 0.1μm to obtain microscopic structure image data;

[0013] Step S13: Obtain a plant component database; extract raw material components from the plant component database to obtain plant component distribution data;

[0014] Step S14: Perform principal component analysis to reduce the dimensionality of the raw material hyperspectral image data, extract the top 10 principal components, and measure the spectral feature vector; use a preset convolutional neural network to extract texture features from the microstructure image data to obtain morphological feature vectors;

[0015] Step S15: Align the spectral feature vector, morphological feature vector, and plant component distribution data across modes and match them according to spatial coordinates to obtain a multi-source plant feature dataset.

[0016] This invention utilizes a hyperspectral imager and a high-resolution microscopic imaging system to acquire multi-scale, multi-modal data from plant materials, obtaining rich spectral and microscopic tissue structure information, thereby achieving comprehensive capture of plant material characteristics. The acquired hyperspectral image data covers the 400-2500nm wavelength band, combined with a fine-grained wavelength resolution of 5nm, ensuring high precision and high-dimensional feature representation of the spectral data. Microscopic structural images capture cell-level tissue morphology with a pixel precision of 0.1μm, providing detailed structural information on the microscopic texture of plant tissues. The introduction of a plant component database provides standardized component distribution reference data, laying a spatial benchmark for subsequent data fusion. In terms of data processing, principal component analysis (PCA) is used to reduce the dimensionality of the hyperspectral data, selecting the top 10 most representative principal components. This significantly reduces data dimensionality while retaining most of the spectral variation information, achieving efficient spectral feature vector construction. Microscopic structural images are then processed using a convolutional neural network (CNN) deep learning model to extract texture features, effectively capturing complex morphological features. Finally, cross-modal data alignment technology is employed to accurately match and fuse spectral feature vectors, morphological feature vectors, and compositional distribution data using a spatial coordinate system, forming a unified multi-source feature dataset. This fusion process not only ensures spatial consistency among different data modalities but also provides a solid multi-dimensional input foundation for subsequent 3D feature space construction and accurate model training.

[0017] Preferably, step S2, which involves constructing a three-dimensional feature space matrix of plants based on a multi-source plant feature dataset, includes:

[0018] The spectral feature vectors are timestamped with plant growth cycle data to obtain plant time dimension encoded data.

[0019] The vascular bundle distribution coordinates of morphological feature vectors were analyzed and spatial topological mapping was performed to obtain plant spatial topological coding data.

[0020] Plant temporal dimension encoded data, plant spatial topological encoded data, and plant component distribution data are concatenated using tensors and a feature matrix is ​​constructed to obtain a three-dimensional feature space matrix of plants.

[0021] This invention effectively introduces temporal dimension information by timestamping spectral feature vectors with plant growth cycle data, achieving a dynamic temporal expression of spectral data and forming plant temporal dimension encoded data. This time stamp not only captures the spectral changes of plants at different growth stages but also provides a quantitative basis for subsequent temporal analysis and dynamic prediction. Simultaneously, precise analysis of vascular bundle distribution coordinates is performed on morphological feature vectors, and a spatial topological mapping method is used to construct the spatial adjacency relationships of tissue structures, forming plant spatial topological encoded data. This encoding system enhances the structured description capability of the complexity of plant micromorphology by expressing the spatial connectivity and distribution patterns of tissue structures. Subsequently, the temporal dimension encoded data, spatial topological encoded data, and plant component distribution data are concatenated at the tensor level to construct a three-dimensional feature space matrix with multi-dimensional information fusion characteristics. This matrix not only contains multimodal information on temporal dynamics, physiological structure, and spatial distribution of components but also achieves efficient data storage and computational adaptation through tensor structures, facilitating high-dimensional feature extraction and predictive analysis in subsequent models. Overall, this process achieves deep integration of plant characteristic data across time, space, and composition dimensions, providing a multidimensional and structured data foundation for comprehensive modeling of plant growth dynamics and composition distribution, and significantly improving the completeness of data representation and the accuracy of prediction models.

[0022] Preferably, step S2, which involves establishing a model for the three-dimensional feature space matrix of the plant, includes:

[0023] Gradient boosting tree analysis based on the three-dimensional feature space matrix of plants yields plant spectral-morphological feature vectors.

[0024] Image feature value analysis was performed on the spectral-morphological feature vectors of plants, and gray value sampling was performed to obtain plant image feature analysis data; the content of active ingredients was predicted from the plant image feature analysis data, and the data units were unified to obtain the predicted values ​​of flavonoid-saponin content;

[0025] The component distribution of the object space topological coding data is processed by random forest classification to obtain the preprocessed plant component data for model construction; the extraction part priority is compared by the preprocessed plant component data for model construction to obtain the plant extraction part priority data.

[0026] The priority data of plant extraction parts and the predicted values ​​of flavonoid-saponin content were used to predict the process parameters using a model, and gradient regression analysis was performed to obtain the plant process parameter prediction model.

[0027] This invention employs the Gradient Boosting Tree (GBT) algorithm to deeply analyze the three-dimensional feature space matrix of plants, effectively fusing spectral and morphological features to generate highly recognizable plant spectral-morphological feature vectors. This process, through iterative weighted optimization, enhances the feature vectors' ability to capture complex nonlinear relationships. Subsequently, image feature value analysis is performed on the obtained feature vectors, and image pixel intensity is quantified using grayscale sampling technology to form plant image feature analysis data, thereby achieving a numerical expression of micro-tissue texture and morphological changes. Based on this, a regression model is used to predict the content of active ingredients in the image feature analysis data. After data unit standardization, predicted values ​​of flavonoid-saponin content are obtained, realizing the transformation from multidimensional features to quantitative component indicators. Simultaneously, a random forest classification algorithm is used to classify the component distribution information in the plant spatial topological coding data, extracting key preprocessed feature data to provide clear component spatial structure information for subsequent model construction. These preprocessed data are then compared based on a priority ranking method to systematically select extraction sites with abundant content and key distribution, forming priority data for plant extraction sites. Finally, using the priority data of extraction sites and the predicted values ​​of flavonoid-saponin content as inputs, and combining gradient regression analysis, a structured and accurate plant process parameter prediction model was formed. This model not only integrates the complex relationships of multi-source and multi-dimensional data, but also achieves effective connection between component prediction and process parameter optimization through multi-stage machine learning methods. From a data perspective, it greatly improves the accuracy and robustness of the prediction model, providing solid data support for the precise extraction process design of plant active ingredients.

[0028] Preferably, step S3 includes the following steps:

[0029] Step S31: Acquire real-time plant images;

[0030] Step S32: Input the real-time plant acquisition image into the plant process parameter prediction model to predict the composition and obtain the spatial distribution prediction map of plant composition; predict the optimal extraction location from the spatial distribution prediction map of plant composition to obtain the optimal location instruction.

[0031] Step S33: Generate the spatial distribution prediction map of plant components and the optimal instruction execution process parameters for each part, and conduct a small-batch extraction experiment to verify the prediction yield data of plants.

[0032] Step S34: Perform reinforcement learning on the plant predicted yield data and map the model weights to obtain parameter prediction accuracy compensation; iterate the process parameters with the parameter prediction accuracy compensation and the plant predicted yield data to obtain optimized process parameter data.

[0033] This invention provides timely and rich visual information data for subsequent component prediction by acquiring high-resolution real-time plant images, ensuring data timeliness and spatial integrity. Step S32 inputs the image into a pre-constructed plant process parameter prediction model. This model, based on multimodal feature fusion and machine learning algorithms, can accurately infer the spatial distribution of active components within the plant, generating a multidimensional component spatial distribution prediction map. This map is further combined with spatial analysis and optimization algorithms to predict the optimal extraction site, outputting optimal instructions for sites with clear spatial positioning and priority to guide the extraction process. Step S33 fuses the component prediction map with the extraction site instructions, automatically formulating corresponding process parameter schemes based on the parameter generation mechanism. Subsequently, small-batch experiments are used to empirically verify the scheme, obtaining plant prediction yield data. This data quantitatively describes the actual performance of extraction efficiency and component retention. Step S34, based on a reinforcement learning framework, uses the prediction yield data as feedback to construct a state-action-reward mapping relationship, dynamically adjusting the model weights to form a compensation mechanism for parameter prediction accuracy. This compensation mechanism combines prediction error and actual yield data, updating process parameters through iterative optimization algorithms to achieve closed-loop optimization of the prediction model and process scheme. Overall, this technology realizes a multi-level data-driven process optimization process, from real-time data acquisition, intelligent prediction, experimental verification to model adaptive iteration, which significantly improves the dynamic response capability and prediction accuracy of process parameters, and provides a scientific data foundation and technical support for the precise control of plant extraction processes.

[0034] Preferably, step S33 includes the following steps:

[0035] Step S331: Query the spatial distribution prediction map of plant components based on the plant component database, and match the solvent to obtain the optimal solvent type data for the plant.

[0036] Step S332: Perform temperature sensitivity analysis based on the predicted flavonoid-saponin content, and construct the extraction curve by extracting at 40℃ for 120 min to obtain the plant extraction temperature-attenuation curve;

[0037] Step S333: Perform texture analysis on the tissue hardness features of real-time plant acquisition images to obtain mechanical breakage strength parameters;

[0038] Step S334: Validate the optimal solvent type data, optimal part instructions, plant extraction temperature-attenuation curve, and mechanical crushing strength parameters through small-batch extraction experiments to obtain plant predicted yield data.

[0039] This invention utilizes a plant component database to query spatial distribution prediction maps. Combining solvent properties and component solubility characteristics from the database, a matching algorithm is used to systematically select the most suitable solvent type for the target active ingredient, generating optimal solvent type data for plants and achieving a quantitative correlation between components and solvents. Step S332, based on the predicted flavonoid-saponin content, conducts temperature sensitivity analysis. Through fitting time-temperature gradient experimental data, a temperature-decay curve is constructed for extraction at 40℃ for 120 minutes, quantitatively describing the stability and decay trend of the active ingredient with temperature and time changes, providing a dynamic function model for temperature parameter optimization. Step S333 performs texture analysis on the tissue hardness characteristics in real-time acquired plant images. Image processing algorithms are used to extract the gray-level co-occurrence matrix and texture statistical features, deriving mechanical fragmentation strength parameters that reflect the mechanical properties of plant cell walls, providing a quantitative basis for fragmentation process parameters. Finally, step S334 integrates the above optimal solvent type, site selection instructions, extraction temperature-decay curve, and mechanical fragmentation strength parameters to guide small-batch extraction experiments. The experimental results are fed back to form plant predicted yield data. This data integrates the synergistic effects of multiple factors, including solvent chemical compatibility, temperature-dependent dynamic response, and tissue mechanical properties, ensuring the scientific control and result verification of the extraction process. Overall, this step, through the fusion analysis of multidimensional data and experimental verification, systematically reveals the intrinsic relationships between key parameters affecting plant extraction efficiency, laying a solid data foundation for the quantitative and personalized optimization of process parameters.

[0040] Preferably, step S34 includes the following steps:

[0041] Step S341: Obtain actual plant yield data;

[0042] Step S342: Perform difference calculation between the actual plant yield data and the predicted plant yield data to obtain the model difference error matrix;

[0043] Step S343: Perform reinforcement learning on the plant prediction yield data and map the model weights to obtain parameter prediction accuracy compensation;

[0044] Step S344: Iterate the process parameters by combining the parameter prediction accuracy compensation with the model difference error matrix to obtain optimized process parameter data.

[0045] This invention uses real-time acquisition of actual plant extraction yield data as a true performance indicator of the process, ensuring the timeliness and accuracy of the data. Step S342 performs difference calculations between the actual yield data and the predicted yield data to form a model difference error matrix. This matrix accurately quantifies the deviation between the model's prediction results and the actual process output, revealing the distribution characteristics of the model's prediction error under different process conditions. Subsequently, step S343, based on a reinforcement learning algorithm, uses the difference error matrix as a feedback signal and adjusts the model weights through a state-action-reward mechanism to achieve adaptive compensation for the prediction parameters. This process establishes a mapping relationship between the model's predicted data and actual observations, enhancing the model's generalization ability and robustness under varying process environments. Finally, step S344 fuses the parameter prediction accuracy compensation results with the error matrix information and updates the process parameters using an iterative optimization algorithm, generating optimized process parameter data. This iterative process continuously corrects prediction deviations, constantly approaching the actual optimal process state, thus improving the accuracy and stability of the process parameters. Overall, this step establishes a closed-loop optimization system from actual process feedback to the self-adjustment of the prediction model, realizing precise optimization of process parameters based on data-driven approaches, and significantly improving the controllability of the extraction process and the adaptability of the prediction model.

[0046] Preferably, step S4 includes the following steps:

[0047] Step S41: Modularly package the optimized process parameter data and construct an active ingredient enrichment heat map;

[0048] Step S42: Overlay the active ingredient enrichment heat map with the optimal extraction site and provide process parameter prompts to obtain a plant extraction process database;

[0049] Step S43: Construct a decision report for the plant extraction process database and generate a plant extraction process decision report.

[0050] This invention optimizes process parameter data through structured and modular processing, facilitating rapid retrieval and flexible combination, thus improving data reusability and management efficiency. Based on this, an active ingredient enrichment heatmap is constructed. By numerically mapping the concentration of active ingredients at different locations within a spatial coordinate system, an intuitive two-dimensional or three-dimensional visualization is formed, enhancing the spatial perception of the distribution characteristics of active ingredients. Step S42 spatially overlays the active ingredient enrichment heatmap with information on the optimal extraction site. Through multi-layer data fusion, a plant extraction process database is formed. This database system integrates the spatial distribution of active ingredients, extraction priorities, and corresponding process parameter prompts, achieving integrated management of multi-source information and intelligent reasoning support. This process not only ensures data consistency and integrity but also improves the database's adaptability and scalability to complex process conditions. Step S43, based on the plant extraction process database, uses a rule engine and data-driven algorithms to construct a decision report, generating a structured plant extraction process decision report. The report content covers the optimal extraction scheme, process parameter configuration, and expected benefit indicators, supporting scientific decision-making and standardized execution of the production process. Overall, this technology, through the structured processing, fusion, and intelligent analysis of multi-level data, achieves precise visualization of active ingredient information and a digital closed loop for process decision-making, greatly improving the transparency, scientific rigor, and efficiency of plant extraction processes.

[0051] Preferably, step S41 includes the following steps:

[0052] Step S411: Iteratively optimize the plant process parameter prediction model using the optimized process parameter data to obtain the plant process optimization prediction model;

[0053] Step S412: Obtain input data, which includes raw material images, composition distribution maps, and yield data;

[0054] Step S413: Based on the input data, the plant process optimization prediction model is encapsulated into modules. When the input data is raw image data, the optimal part data is obtained; when the input data is a component distribution map, the solvent-temperature-crushing parameters are obtained; when the input data is yield data, the efficacy potential score data is obtained.

[0055] Step S414: Perform three-dimensional reconstruction of the site selection data, solvent-temperature-crushing parameters and efficacy potential score data to obtain an active ingredient enrichment heat map.

[0056] This invention utilizes the latest optimized process parameter data to iteratively train and adjust the original process parameter prediction model, effectively improving the model's adaptability to the diversity of plant raw materials and the complexity of processes, and generating a more accurate process optimization model. In step S412, the system collects multi-dimensional inputs including raw material images, component distribution maps, and yield data. These data cover visual information of plant raw materials, spatial distribution of active ingredients, and actual extraction efficiency, ensuring the diversity and comprehensiveness of the input data. Step S413 designs a modular processing path for different types of input data: when the input is raw image data, the model outputs optimal site data through image feature extraction and pattern recognition technology, achieving spatial positioning and priority determination of extraction sites; when the input is a component distribution map, it combines chemical properties and thermodynamic parameters to predict solvent selection, temperature control, and mechanical crushing parameters, achieving optimized configuration of key process variables; when the input is yield data, it calculates efficacy potential score data through regression and scoring algorithms, providing quantitative evaluation indicators for process effectiveness. Finally, step S414 performs three-dimensional reconstruction of the aforementioned multi-source output data to construct an active ingredient enrichment heatmap. This heatmap visually displays the concentration distribution of active ingredients in different parts of the plant and the comprehensive influence of process parameters in three-dimensional space, enhancing the spatial expressiveness and decision support function of the data. Overall, this technical step achieves dynamic updating of the process optimization model based on multimodal input and multi-dimensional data fusion, providing scientific, precise, and dynamically adaptable parameter support for plant extraction processes, significantly improving the targeting of process design and the accuracy of execution.

[0057] This specification provides an image processing-based plant extraction parameter setting system for performing the above-described image processing-based plant extraction parameter setting method. The image processing-based plant extraction parameter setting system includes:

[0058] The multimodal feature construction module is used to deploy multimodal data acquisition sensors on plant raw materials and perform multi-scale data fusion to construct a multi-source feature dataset of plants.

[0059] The 3D feature modeling and parameter prediction module is used to construct a 3D feature space matrix of plants based on multi-source plant feature datasets; and to establish a model of the 3D feature space matrix of plants to obtain a plant process parameter prediction model.

[0060] The real-time image prediction and reinforcement learning optimization module is used to acquire real-time plant images; the real-time plant images are used as input to the plant process parameter prediction model to predict components, resulting in a spatial distribution prediction map of plant components; the spatial distribution prediction map of plant components is validated through small-batch extraction experiments to obtain plant prediction yield data; reinforcement learning is applied to the plant prediction yield data, and the relational mapping model weights are used to obtain parameter prediction accuracy compensation; the parameter prediction accuracy compensation and the plant prediction yield data are used to iterate the process parameters to obtain optimized process parameter data.

[0061] The process parameter encapsulation and decision visualization module is used to modularly encapsulate optimized process parameter data and construct an active ingredient enrichment heat map; it overlays the active ingredient enrichment heat map with the optimal extraction site and provides process parameter prompts to obtain a plant extraction process decision report.

[0062] The beneficial effects of this invention lie in its ability to comprehensively acquire and express high-dimensional characteristics of plant raw materials through multimodal data acquisition and multi-scale fusion, thereby constructing a three-dimensional feature space matrix covering spectral, morphological, and component spatial distribution. This provides a rich and structured data foundation for the accurate prediction of plant processing parameters. During data processing, real-time plant images are inferred through deep learning using a process parameter prediction model to generate a component spatial distribution prediction map. Combined with prediction yield data obtained from small-batch experimental verification, the model weights are dynamically adjusted using a reinforcement learning algorithm, achieving continuous optimization and adaptive compensation of parameter prediction accuracy. The optimized process parameter data is further modularly packaged into standardized inputs. A heatmap of active ingredient enrichment is constructed using spatial overlay technology, accurately revealing the spatial distribution characteristics of target components in different tissues. Combined with a process parameter prompting system, this effectively guides the selection of extraction sites and the formulation of process plans, ultimately generating a plant extraction process decision report. This report, based on multi-source fusion data and iterative optimization results, provides scientific and quantitative decision support for process execution, reducing the blindness and experience-based dependence of traditional process parameter setting and improving the accuracy and controllability of the extraction process. Attached Figure Description

[0063] Figure 1 A schematic diagram illustrating the steps of a plant extraction parameter setting method based on image processing;

[0064] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S4.

[0065] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0066] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0067] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0068] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0069] To achieve the above objectives, please refer to Figures 1 to 2 A method for setting plant extraction parameters based on image processing, the method comprising the following steps:

[0070] Step S1: Deploy multimodal data acquisition sensors on plant materials and perform multi-scale data fusion to construct a multi-source feature dataset of plants;

[0071] Step S2: Construct a three-dimensional feature space matrix of plants based on a multi-source feature dataset; establish a model for the three-dimensional feature space matrix of plants to obtain a prediction model for plant process parameters;

[0072] Step S3: Acquire real-time plant images; input the real-time plant images into the plant process parameter prediction model to predict plant components, and obtain a spatial distribution prediction map of plant components; perform small-batch extraction experiments to verify the spatial distribution prediction map of plant components, and obtain plant prediction yield data; perform reinforcement learning on the plant prediction yield data and map the model weights to obtain parameter prediction accuracy compensation; iterate the parameter prediction accuracy compensation and the plant prediction yield data to obtain optimized process parameter data.

[0073] Step S4: Modularly package the optimized process parameter data and construct an active ingredient enrichment heat map; overlay the active ingredient enrichment heat map with the optimal extraction sites and provide process parameter suggestions to obtain a plant extraction process decision report.

[0074] In this embodiment of the invention, reference Figure 1 The diagram shown is a flowchart illustrating the steps of a plant extraction parameter setting method based on image processing according to the present invention. In this example, the plant extraction parameter setting method based on image processing includes the following steps:

[0075] Preferably, step S1 includes the following steps:

[0076] Step S11: Use a hyperspectral imager with a wavelength range of 400-2500nm and a resolution of 5nm to scan the plant raw materials and collect hyperspectral image data of the raw materials.

[0077] Step S12: Acquire images of plant tissue cell structures using a microscopic imaging system with a magnification of 40-1000X and a pixel accuracy of 0.1μm to obtain microscopic structure image data;

[0078] Step S13: Obtain a plant component database; extract raw material components from the plant component database to obtain plant component distribution data;

[0079] Step S14: Perform principal component analysis to reduce the dimensionality of the raw material hyperspectral image data, extract the top 10 principal components, and measure the spectral feature vector; use a preset convolutional neural network to extract texture features from the microstructure image data to obtain morphological feature vectors;

[0080] Step S15: Align the spectral feature vector, morphological feature vector, and plant component distribution data across modes and match them according to spatial coordinates to obtain a multi-source plant feature dataset.

[0081] In this embodiment of the invention, a spectral feature vector matrix of the raw materials is constructed based on the top 10 principal components extracted in step S14. This matrix retains the spectral variation features with the highest information content in the hyperspectral band and eliminates interference caused by changes in the acquisition environment through normalization and standardization. The morphological feature vector is extracted from the image's texture, cell arrangement, shape contour, and edge direction at multiple receptive field scales using a convolutional neural network, forming a high-dimensional representation at the structural level. Meanwhile, the plant component distribution data is derived from mappings of existing databases, typically using quantitative indicators at the regional or cellular scale to represent the probability of presence or concentration gradient of specific active ingredients. To achieve alignment of the three types of data, their spatial coordinate systems must first be uniformly transformed. A registration algorithm based on multiple interpolation is used to map the hyperspectral pixel grid, microscopic image resolution units, and component distribution labels to a standard spatial grid at the same scale. After coordinate registration, a pairing method based on maximum mutual information is further used for cross-modal feature alignment to ensure that each spatial point simultaneously possesses spectral, structural, and component attributes. The final constructed plant multi-source feature dataset is represented in the form of a multidimensional tensor, where each tensor unit contains a composite feature vector corresponding to a spatial location. This vector includes measured principal component spectral lines, microscopic texture descriptors, and standard component labels, forming a unified, multimodal fusion basic data structure that provides a highly consistent and complementary information foundation for subsequent feature modeling and prediction tasks.

[0082] Preferably, step S2, which involves constructing a three-dimensional feature space matrix of plants based on a multi-source plant feature dataset, includes:

[0083] The spectral feature vectors are timestamped with plant growth cycle data to obtain plant time dimension encoded data.

[0084] The vascular bundle distribution coordinates of morphological feature vectors were analyzed and spatial topological mapping was performed to obtain plant spatial topological coding data.

[0085] Plant temporal dimension encoded data, plant spatial topological encoded data, and plant component distribution data are concatenated using tensors and a feature matrix is ​​constructed to obtain a three-dimensional feature space matrix of plants.

[0086] In this embodiment of the invention, for the spectral feature vector, a timestamp related to the plant growth cycle is embedded in its original data structure. This timestamp uses the actual collection date, growth stage, accumulated temperature, or leaf age as time index variables. The time information is numerically processed through periodic function encoding (such as sine / cosine encoding or position encoding) to form a continuous and calculable time dimension feature, thereby enabling the original spectral data to express dynamic changes over time, thus constructing plant time dimension encoded data. Subsequently, the morphological feature vectors extracted from the microscopic image are spatially located. Image segmentation and morphological annotation techniques are used to extract the contours of tissue structures such as vascular bundles. A structural center distribution map is constructed based on the centroid coordinates and local geometric relationships. Then, the connection relationships between structural units are modeled through topological structure mapping methods (such as Voronoi diagrams or Delaunay triangular meshes), and their spatial adjacency matrices, edge weight relationships, or topological kernel function expressions are extracted, ultimately generating plant spatial topological encoded data. After completing the temporal and spatial encoding, the two types of data need to be aligned with the plant component distribution data in multiple dimensions. First, a spatial registration algorithm is used to map the three types of data to a unified coordinate system. While ensuring consistency between the physical space and the information space, the temporal encoding vector, topological encoding vector, and component concentration data at the corresponding locations are retrieved using spatial location as the key index. Next, a tensor concatenation operation is used to connect the three types of data into the same vector group along the feature dimension, and then expands them into a regular three-dimensional tensor structure according to the spatial structure. The first dimension corresponds to the spatial location index, the second dimension corresponds to the dimension of the concatenated multimodal feature vector, and the third dimension can represent the stacked form of multi-time point samples or multiple raw material batches as needed. To ensure the computability of the tensor in subsequent modeling, the concatenation result needs to be preprocessed, including standardization, zero-mean normalization, and missing value imputation. Finally, a three-dimensional feature space matrix of plants with temporal semantics, spatial topological structure, and component distribution logic is formed, providing a unified data representation basis for feature extraction and predictive modeling.

[0087] Preferably, step S2, which involves establishing a model for the three-dimensional feature space matrix of the plant, includes:

[0088] Gradient boosting tree analysis based on the three-dimensional feature space matrix of plants yields plant spectral-morphological feature vectors.

[0089] Image feature value analysis was performed on the spectral-morphological feature vectors of plants, and gray value sampling was performed to obtain plant image feature analysis data; the content of active ingredients was predicted from the plant image feature analysis data, and the data units were unified to obtain the predicted values ​​of flavonoid-saponin content;

[0090] The component distribution of the object space topological coding data is processed by random forest classification to obtain the preprocessed plant component data for model construction; the extraction part priority is compared by the preprocessed plant component data for model construction to obtain the plant extraction part priority data.

[0091] The priority data of plant extraction parts and the predicted values ​​of flavonoid-saponin content were used to predict the process parameters using a model, and gradient regression analysis was performed to obtain the plant process parameter prediction model.

[0092] In this embodiment of the invention, a three-dimensional feature space matrix of plants is used as input. A Gradient Boosting Decision Tree (GBDT) model is employed to perform nonlinear feature importance ranking and combination on high-dimensional spectral features and morphological topological data, uncovering strong correlations between these features and target component variables. This results in a fused, representative plant spectral-morphological feature vector. This vector serves as the basis for further extraction of image structural information. It is input into an image feature value analysis module for feature map generation and grayscale mapping. Texture analysis methods such as Gray-Level Co-occurrence Matrix (GLCM) are used to quantify features such as image brightness, contrast, and entropy, obtaining plant image feature analysis data. Subsequently, a regression prediction model is constructed based on this feature data. Combined with experimental standard curves or database annotation values, the contents of flavonoids and saponins are predicted using multiple linear regression, partial least squares, or lightweight neural networks. The predicted values ​​are then subjected to unit conversion and concentration normalization to obtain component prediction indicators under a unified standard. In parallel, plant spatial topological encoding data is input into a random forest classifier. Based on the spatial distribution patterns of components in the training samples, the classifier learns the ability of different structural regions to distinguish component accumulation, outputting a component enrichment potential label for each topological region as preprocessing data for model building of plant components. These regions are then compared laterally using a multi-region priority evaluation function, calculating comprehensive scores in terms of extraction efficiency, component enrichment rate, and stability to obtain priority data for plant extraction parts. Finally, the priority data is concatenated with the predicted flavonoid-saponin content to construct an input vector. Gradient regression methods (such as XGBoost regression or multilayer perceptron optimized by gradient descent) are used to model the output target process parameters such as extraction temperature, time, pH value, and solvent ratio, completing the training and expression structure encapsulation of the plant process parameter prediction model, forming an end-to-end modeling path.

[0093] Preferably, step S3 includes the following steps:

[0094] Step S31: Acquire real-time plant images;

[0095] Step S32: Input the real-time plant acquisition image into the plant process parameter prediction model to predict the composition and obtain the spatial distribution prediction map of plant composition; predict the optimal extraction location from the spatial distribution prediction map of plant composition to obtain the optimal location instruction.

[0096] Step S33: Generate the spatial distribution prediction map of plant components and the optimal instruction execution process parameters for each part, and conduct a small-batch extraction experiment to verify the prediction yield data of plants.

[0097] Step S34: Perform reinforcement learning on the plant predicted yield data and map the model weights to obtain parameter prediction accuracy compensation; iterate the process parameters with the parameter prediction accuracy compensation and the plant predicted yield data to obtain optimized process parameter data.

[0098] In this embodiment of the invention, a high-resolution imaging device is used to acquire real-time plant images. During image acquisition, location information and raw material batch identifiers are recorded simultaneously to ensure data consistency with information in the existing three-dimensional feature space matrix. After image acquisition, in step S32, the image is input into a pre-constructed plant process parameter prediction model. The model, based on a deep convolutional neural network architecture, combines spectral feature transfer learning and topology perception mechanisms to predict the component distribution of each spatial pixel or region in the image, outputting a plant component spatial distribution prediction map. This map is a two-dimensional or three-dimensional spatial mapping map, where each unit contains corresponding component concentration or presence probability data. Subsequently, an optimal extraction region identification algorithm based on component concentration gradient and spatial clustering is introduced. Combining boundary saliency calculation and cluster analysis, the region with the highest target component enrichment and strongest distribution stability is selected to generate the optimal site instruction, and structural coordinates and region priority labels are output. In step S33, a preset micro-extraction experiment is executed according to the prediction map and optimal instruction parameters to extract samples from the specified region, record the actual yield data, and generate a plant predicted yield dataset. After proceeding to step S34, error analysis is performed between the predicted yield data and the model's original prediction results. An error feedback loop is constructed using reinforcement learning strategies—such as policy gradient-based or Q-learning methods—and the weights of key layers controlling process parameter outputs in the model are fine-tuned to form a parameter prediction accuracy compensation vector. This compensation vector, serving as the basis for dynamic weight adjustment, is input along with the original predicted yield data into the process parameter optimization module. This module performs iterative parameter calculations based on a multi-objective loss function, enhancing the model's generalization performance while maximizing the yield. Ultimately, it outputs optimized process parameter data with a standardized structure and superior predictive ability, providing a newer foundation for large-scale extraction of process parameters.

[0099] Preferably, step S33 includes the following steps:

[0100] Step S331: Query the spatial distribution prediction map of plant components based on the plant component database, and match the solvent to obtain the optimal solvent type data for the plant.

[0101] Step S332: Perform temperature sensitivity analysis based on the predicted flavonoid-saponin content, and construct the extraction curve by extracting at 40℃ for 120 min to obtain the plant extraction temperature-attenuation curve;

[0102] Step S333: Perform texture analysis on the tissue hardness features of real-time plant acquisition images to obtain mechanical breakage strength parameters;

[0103] Step S334: Validate the optimal solvent type data, optimal part instructions, plant extraction temperature-attenuation curve, and mechanical crushing strength parameters through small-batch extraction experiments to obtain plant predicted yield data.

[0104] In this embodiment of the invention, a plant component database is used as a knowledge base to perform data query and matching operations on the spatial distribution prediction map of plant components. The database stores the optimal solvent type and its solubility characteristics corresponding to different active components. Through a matching algorithm of spatial coordinates and component labels, combined with multi-dimensional data such as solvent solubility parameters, polarity index, and diffusion coefficient, the optimal solvent type of each active component in the prediction map is accurately inferred, forming plant optimal solvent type data. Secondly, in step S332, based on the predicted value of flavonoid-saponin content, temperature sensitivity analysis is carried out. The data input includes the predicted component concentration and its chemical stability parameters as a function of temperature. By establishing a time-concentration response curve under a constant temperature extraction condition of 40℃ for 120 minutes, a nonlinear fitting method (such as an exponential decay model or a double exponential model) is used to construct a plant extraction temperature-decay curve, reflecting the extraction efficiency and decay trend of components over time. In step S333, tissue stiffness features are extracted from the real-time acquired plant images. Texture analysis techniques (including gray-level co-occurrence matrix and local binary mode methods) are used to quantify the roughness and toughness of the cell wall structure, thereby deriving the mechanical breakage strength parameter. This parameter reflects the physical resistance of the plant tissue and provides a mechanical basis for subsequent extraction process design. Finally, in step S334, the optimal solvent type data, optimal site instructions, extraction temperature-attenuation curve, and mechanical breakage strength parameter generated above are integrated as multidimensional input variables to conduct small-batch extraction experiments. The extraction efficiency, component yield, and extraction stability are monitored during the experiment, and real-time yield data is collected to form a plant predicted yield dataset. This process uses multidimensional arrays and time series matrices to store experimental data to support subsequent model validation and optimization iterations.

[0105] Preferably, step S34 includes the following steps:

[0106] Step S341: Obtain actual plant yield data;

[0107] Step S342: Perform difference calculation between the actual plant yield data and the predicted plant yield data to obtain the model difference error matrix;

[0108] Step S343: Perform reinforcement learning on the plant prediction yield data and map the model weights to obtain parameter prediction accuracy compensation;

[0109] Step S344: Iterate the process parameters by combining the parameter prediction accuracy compensation with the model difference error matrix to obtain optimized process parameter data.

[0110] In this embodiment of the invention, actual yield data from small-batch experiments of plant extraction processes are collected. The data is recorded in a multi-dimensional matrix, including sample number, extraction time point, extraction conditions, and corresponding yield value, ensuring the temporal integrity and spatial identification accuracy of the data. Subsequently, in step S342, the actual experimental yield data and the plant predicted yield data output by the previous prediction model are calculated element-wise to form a model difference error matrix. This matrix reflects the deviation distribution between the model prediction and the actual observation, supporting the analysis of the spatial and temporal distribution characteristics of the error. In step S343, a reinforcement learning framework is constructed based on the difference error matrix. Using error feedback as the reward signal, the key weight parameters in the prediction model are dynamically adjusted using policy gradient or Q-learning algorithms. This process, through a state-action-reward mechanism, gradually optimizes the model weight mapping function, achieving adaptive compensation for parameter prediction accuracy and enhancing the model's robustness to sample diversity and environmental fluctuations. Finally, step S344 inputs the compensated prediction accuracy parameters and the difference error matrix into the process parameter iteration module. Using numerical calculation methods such as gradient descent or Bayesian optimization, multiple rounds of iterative optimization are performed on the process parameter space to correct key variables such as extraction temperature, time, and solvent ratio, forming a structured optimized process parameter dataset. This dataset is stored in tensor form, supporting subsequent process execution and model iteration updates, completing the experimental data-driven closed-loop optimization process.

[0111] As an example of the present invention, reference is made to Figure 2 As shown, step S4 in this example includes:

[0112] Step S41: Modularly package the optimized process parameter data and construct an active ingredient enrichment heat map;

[0113] Step S42: Overlay the active ingredient enrichment heat map with the optimal extraction site and provide process parameter prompts to obtain a plant extraction process database;

[0114] Step S43: Construct a decision report for the plant extraction process database and generate a plant extraction process decision report.

[0115] In this embodiment of the invention, the optimized process parameter data undergoes multi-dimensional vectorization processing, including parameters such as temperature, time, solvent ratio, and ultrasonic intensity. A modular data encapsulation framework is used to hierarchically archive these parameters according to batch, acquisition time, and process type, forming standardized parameter modules with metadata tags. Subsequently, based on the component distribution information output by experimental measurements and prediction models, an active ingredient enrichment heatmap is generated using kernel density estimation and spatial interpolation algorithms. This heatmap is expressed in two-dimensional or three-dimensional raster data form, quantifying the relative intensity of active ingredient concentration within each grid cell in a spatial coordinate system. After normalization processing, a continuous numerical heat distribution layer is formed. In step S42, the enrichment heatmap is fused with the previously obtained spatial coordinate data of the optimal extraction site using a spatial overlay algorithm. Multi-scale image registration and weighted fusion techniques are employed to achieve precise spatial positioning and priority division of the extraction area. Simultaneously, a process parameter prompting system is introduced. This system, based on a framework combining a rule engine and a machine learning model, automatically generates parameter adjustment suggestions for different extraction sites based on the heatmap intensity and spatial overlay results, outputting structured prompt information. The integrated data and accompanying prompts form a plant extraction process database, supporting rapid retrieval and subsequent process execution. Finally, in step S43, utilizing the multidimensional process parameter data and heatmap spatial information stored in the database, combined with extraction site priority and parameter prompts, a plant extraction process decision report is automatically generated through data aggregation and visualization templates. The report structure employs a hierarchical data model, including a parameter summary table, heatmap visual presentation, spatial extraction suggestions, and historical comparative analysis. It supports export in multiple formats to meet different application needs, achieving an automated closed loop from data integration to decision support.

[0116] Preferably, step S41 includes the following steps:

[0117] Step S411: Iteratively optimize the plant process parameter prediction model using the optimized process parameter data to obtain the plant process optimization prediction model;

[0118] Step S412: Obtain input data, which includes raw material images, composition distribution maps, and yield data;

[0119] Step S413: Based on the input data, the plant process optimization prediction model is encapsulated into modules. When the input data is raw image data, the optimal part data is obtained; when the input data is a component distribution map, the solvent-temperature-crushing parameters are obtained; when the input data is yield data, the efficacy potential score data is obtained.

[0120] Step S414: Perform three-dimensional reconstruction of the site selection data, solvent-temperature-crushing parameters and efficacy potential score data to obtain an active ingredient enrichment heat map.

[0121] In this embodiment of the invention, optimized process parameter data is used as feedback signals. The plant process parameter prediction model is iteratively trained based on gradient descent or Bayesian optimization methods, adjusting the model parameter weights to dynamically correct the prediction function, ultimately obtaining a plant process optimization prediction model with higher generalization ability and prediction accuracy. In step S412, multi-source input datasets are collected and organized, including high-resolution plant raw material image data, component spatial distribution maps generated by the prediction model, and corresponding small-batch extraction yield data. All three types of data undergo standardization preprocessing to ensure consistent data scale and spatial-temporal alignment. In step S413, the optimized prediction model is encapsulated into functional modules, and independent sub-model interfaces are designed for different types of input data: when the input is a raw plant image, the module outputs site optimization data through image feature extraction and spatial clustering algorithms to characterize the optimal tissue location for active ingredient enrichment; when the input is a component distribution map, the model uses multivariate regression or deep learning decoding networks to calculate the corresponding solvent type, extraction temperature, and mechanical crushing parameters to generate key variables for the extraction process; when the input is yield data, the extraction effect is evaluated through classification or regression models, and efficacy potential score data is output to measure the actual performance of the extraction scheme. Finally, in step S414, the output data of the above three types of modules are mapped to a unified three-dimensional spatial coordinate system, and spatial interpolation and tensor reconstruction techniques are used to fuse the data to generate an active ingredient enrichment heatmap containing spatial location, component enrichment intensity, and process indicators. This three-dimensional heatmap is stored in raster data form, supporting spatial visualization and subsequent decision analysis, realizing a complete data flow from multi-source data fusion to spatial feature expression.

[0122] Of particular importance, step S42 includes the following steps:

[0123] Step S421: Perform spatial coordinate analysis on the active ingredient enrichment heatmap and extract the coordinate point set of the concentration peak region to obtain the ingredient enrichment coordinate data;

[0124] Step S422: Call the microscopic structure image for biological anatomy recognition, including edge detection of vascular bundles to obtain tissue structure boundary data, and morphological analysis of cell wall thickness to obtain tissue stiffness distribution data;

[0125] Step S423: Spatial registration and overlay of component enrichment coordinate data and tissue structure boundary data, and matching of extraction sites using tissue hardness distribution data to obtain the optimal extraction site labeling instruction;

[0126] Step S424: Based on the instruction to mark the best extraction site, provide process parameter suggestions to obtain a plant extraction process database.

[0127] In this embodiment of the invention, spatial coordinate analysis is performed on the heatmap of active ingredient enrichment. A peak extraction method based on gradient analysis and region growing algorithm is used to extract significant high-concentration areas from the heatmap and transform them into a set of coordinate points with clear spatial reference, forming ingredient enrichment coordinate data. Subsequently, in step S422, high-resolution microscopic structure image data is called, and image segmentation and feature recognition algorithms are used to complete the fine analysis at the biological anatomical level. Specifically, this includes: applying the Canny edge detection algorithm and structure enhancement filter to extract the edge contours of vascular bundle structures to generate tissue structure boundary data; simultaneously, morphological image processing and texture direction filtering are used to perform pixel-level statistical analysis on cell wall thickness regions to obtain tissue stiffness distribution data. These microscopic data provide morphological basis for subsequent spatial registration and structural adaptation. In step S423, the system performs multi-scale spatial registration of the component enrichment coordinate data from the macroscopic component heatmap with the tissue structure boundary data at the microscopic level. Affine transformation or feature-point-based geometric calibration methods are used to achieve the mapping relationship between images of different resolutions. Local region matching analysis is then performed using tissue stiffness distribution data. A weighted matching function identifies regions in the tissue structure that are both component-rich and have low extraction resistance, ultimately generating an instruction for the optimal extraction site with spatial location annotation. In step S424, this optimal extraction site annotation instruction is used as the driving input. Combined with an existing historical database of extraction processes and parameter templates, structured process parameter prompts are generated through parameter mapping and rule combination. This ultimately forms a plant extraction process database for different plant samples, serving as the foundational data support for subsequent automated extraction or process recommendations. This process achieves end-to-end information fusion from high-dimensional spatial enrichment information to cell structure response characteristics, and through structural analysis and spatial mapping, completes the data-driven construction of the optimal extraction strategy.

[0128] Of particular importance, step S43 includes the following steps:

[0129] Step S431: Extract target entries based on the plant extraction process database to obtain the target process dataset;

[0130] Step S432: Perform triple validation analysis on the target process dataset. The triple validation analysis includes calculating and scoring the relative error between the actual plant yield data and the predicted plant yield data to obtain a yield reliability score; calling the efficacy-component association data from the plant extraction process database to perform causal strength analysis to obtain the efficacy evidence level; and calculating the raw material cost of the plant extraction process database to obtain a cost-benefit index.

[0131] Step S433: Apply the yield reliability score, efficacy evidence level, and cost-effectiveness index to the standard template riverbed to obtain the plant extraction process decision dataset;

[0132] Step S434: Construct a decision report from the plant extraction process decision dataset and generate a plant extraction process decision report.

[0133] In this embodiment of the invention, historical process records matching the target raw material or target component are extracted from a plant extraction process database. A structured target process dataset is obtained by setting keywords, attribute fields for filtering, or using a tag-based indexing mechanism. This dataset contains multi-dimensional data such as extraction method, solvent type, extraction temperature, time parameters, historical yield, efficacy tags, and cost records. Subsequently, in step S432, a triple verification analysis is performed on the target process dataset: First, based on the comparison between actual plant yield data and predicted plant yield data, an error calculation method such as Mean Relative Error (MAPE) or logarithmic loss is used to generate a yield reliability score. Second, the structured correlation data of efficacy and component in the database is called, and the causal strength between key components and target efficacy in the target extraction scheme is quantified through Bayesian causal graph or structural equation model analysis, and a grade label is assigned to form an efficacy evidence level. Third, a multi-factor weighted evaluation is performed on the ratio between raw material cost, equipment operating cost, and unit effective component output to calculate a cost-benefit index, which serves as the basis for economic evaluation. In step S433, the system inputs the evaluation results of the three dimensions mentioned above—yield reliability score, efficacy evidence level, and cost-effectiveness index—into a standardized decision template. It then uses the Analytic Hierarchy Process (AHP) or a weighted linear model to perform unified normalization and weight fusion on the multiple indicators, constructing a clearly structured data decision matrix, thus forming a plant extraction process decision dataset. In step S434, the decision dataset is documented according to the standardized template, including automatically generating summary descriptions, tabulated process parameters, and visual evaluation charts, and marking risk points and recommendation levels. Finally, a well-structured, clearly defined, and easily interpretable plant extraction process decision report is output, providing clear data support and process suggestions for subsequent process optimization, experimental design, or batch extraction. The entire process achieves a closed loop from quantitative extraction of process data to cross-dimensional evaluation and fusion, possessing high information density and structural hierarchy.

[0134] This specification provides an image processing-based plant extraction parameter setting system for performing the above-described image processing-based plant extraction parameter setting method. The image processing-based plant extraction parameter setting system includes:

[0135] The multimodal feature construction module is used to deploy multimodal data acquisition sensors on plant raw materials and perform multi-scale data fusion to construct a multi-source feature dataset of plants.

[0136] The 3D feature modeling and parameter prediction module is used to construct a 3D feature space matrix of plants based on multi-source plant feature datasets; and to establish a model of the 3D feature space matrix of plants to obtain a plant process parameter prediction model.

[0137] The real-time image prediction and reinforcement learning optimization module is used to acquire real-time plant images; the real-time plant images are used as input to the plant process parameter prediction model to predict components, resulting in a spatial distribution prediction map of plant components; the spatial distribution prediction map of plant components is validated through small-batch extraction experiments to obtain plant prediction yield data; reinforcement learning is applied to the plant prediction yield data, and the relational mapping model weights are used to obtain parameter prediction accuracy compensation; the parameter prediction accuracy compensation and the plant prediction yield data are used to iterate the process parameters to obtain optimized process parameter data.

[0138] The process parameter encapsulation and decision visualization module is used to modularly encapsulate optimized process parameter data and construct an active ingredient enrichment heat map; it overlays the active ingredient enrichment heat map with the optimal extraction site and provides process parameter prompts to obtain a plant extraction process decision report.

[0139] The beneficial effects of this invention lie in its ability to comprehensively acquire and express high-dimensional characteristics of plant raw materials through multimodal data acquisition and multi-scale fusion, thereby constructing a three-dimensional feature space matrix covering spectral, morphological, and component spatial distribution. This provides a rich and structured data foundation for the accurate prediction of plant processing parameters. During data processing, real-time plant images are inferred through deep learning using a process parameter prediction model to generate a component spatial distribution prediction map. Combined with prediction yield data obtained from small-batch experimental verification, the model weights are dynamically adjusted using a reinforcement learning algorithm, achieving continuous optimization and adaptive compensation of parameter prediction accuracy. The optimized process parameter data is further modularly packaged into standardized inputs. A heatmap of active ingredient enrichment is constructed using spatial overlay technology, accurately revealing the spatial distribution characteristics of target components in different tissues. Combined with a process parameter prompting system, this effectively guides the selection of extraction sites and the formulation of process plans, ultimately generating a plant extraction process decision report. This report, based on multi-source fusion data and iterative optimization results, provides scientific and quantitative decision support for process execution, reducing the blindness and experience-based dependence of traditional process parameter setting and improving the accuracy and controllability of the extraction process.

[0140] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0141] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. An image processing-based plant extraction parameter setting method, characterized by, Comprise the following steps: Step S1: Deploying multi-modal data acquisition sensors on plant raw materials, and carrying out multi-scale data fusion to construct plant multi-source feature dataset, specifically: through hyperspectral imager and high-resolution microscopic imaging system, multi-scale and multi-modal data acquisition is carried out on plant raw materials, and plant multi-source feature dataset is constructed based on spectral feature vector, morphological feature vector and plant component distribution data; Step S2: Constructing plant three-dimensional feature space matrix based on plant multi-source feature dataset, specifically: obtaining three-dimensional feature space matrix based on plant time dimension coding data, plant space topology coding data and plant component distribution data; model establishment is carried out on plant three-dimensional feature space matrix, and plant process parameter prediction model is obtained; Step S3: Obtain real-time plant acquisition image; the real-time plant acquisition image is taken as input to the plant process parameter prediction model for component prediction to obtain plant component spatial distribution prediction map; the plant component spatial distribution prediction map is subjected to small batch extraction experiment verification to obtain plant prediction yield data; the plant prediction yield data is subjected to reinforcement learning, and the model weight is related to map to obtain parameter prediction accuracy compensation; the parameter prediction accuracy compensation and the plant prediction yield data are subjected to process parameter iteration to obtain optimized process parameter data; Step S4: Iterative training of the optimized process parameter data on the plant process parameter prediction model, and module packaging of the plant process parameter prediction model by using input data, three-dimensional reconstruction according to output data, and obtaining active ingredient enrichment heat map; superimposing the active ingredient enrichment heat map on the best extraction site, and prompting the process parameters to obtain the plant extraction process decision report.

2. The image processing-based plant extraction parameter setting method according to claim 1, characterized in that, Step S1 comprises the following steps: Step S11: Selecting a hyperspectral imager with a wave band range of 400-2500 nm and a resolution of 5 nm to scan the plant raw materials and collect raw material hyperspectral image data; Step S12: Collecting plant tissue cell structure images by a microscopic imaging system with a magnification of 40-1000X and a pixel accuracy of 0.1 μm to obtain microscopic structure image data; Step S13: Obtaining plant component database; extracting raw material components from the plant component database to obtain plant component distribution data; Step S14: Principal component analysis dimension reduction is carried out on the raw material hyperspectral image data, and the first 10 principal components are extracted to measure the spectral feature vector; the texture features of the microscopic structure image data are extracted by using a preset convolutional neural network to obtain the morphological feature vector; Step S15: Cross-modal alignment of the spectral feature vector, morphological feature vector and plant component distribution data is carried out, and spatial coordinate matching is carried out to obtain the plant multi-source feature dataset.

3. The method of claim 1, wherein the method further comprises: In step S2, constructing plant three-dimensional feature space matrix based on plant multi-source feature dataset comprises: Timestamp marking of the spectral feature vector for plant growth cycle to obtain plant time dimension coding data; Vascular bundle distribution coordinate analysis is carried out on the morphological feature vector, and spatial topology mapping is carried out to obtain plant space topology coding data; The plant time dimension encoding data, the plant space topology encoding data and the plant component distribution data are spliced into a tensor, and a feature matrix is constructed to obtain a plant three-dimensional feature space matrix.

4. The method of claim 1, wherein the method further comprises: The model establishment in step S2 on the plant three-dimensional feature space matrix comprises: Gradient boosting tree analysis is performed based on the plant three-dimensional feature space matrix to obtain a plant spectrum-morphological feature vector; Image feature value analysis is performed on the plant spectrum-morphological feature vector, and a gray value is sampled to obtain plant image feature analysis data; the active ingredient content prediction is performed on the plant image feature analysis data, and the data units are unified to obtain the flavonoid-saponin content prediction value; The component distribution of the plant space topology encoding data is processed by random forest classification to obtain model construction plant component preprocessing data; the model construction plant component preprocessing data is extracted and compared in priority to obtain plant extraction site priority data; The plant extraction site priority data and the flavonoid-saponin content prediction value are subjected to process parameter model prediction, and gradient regression analysis is performed to obtain a plant process parameter prediction model.

5. The image processing-based plant extraction parameter setting method according to claim 1, characterized in that, Step S3 comprises the following steps: Step S31: acquiring a real-time plant collection image; Step S32: inputting the real-time plant collection image into the plant process parameter prediction model for component prediction to obtain a plant component space distribution prediction map; the plant component space distribution prediction map is subjected to optimal extraction site prediction to obtain a site optimization instruction; Step S33: generating plant extraction process parameters based on the plant component space distribution prediction map and the site optimization instruction, and performing small-batch extraction experiments to obtain plant predicted yield data; Step S34: performing reinforcement learning on the plant predicted yield data, and mapping the model weight to obtain a parameter prediction accuracy compensation; the parameter prediction accuracy compensation and the plant predicted yield data are subjected to process parameter iteration to obtain optimized process parameter data.

6. The image processing-based plant extraction parameter setting method according to claim 5, wherein, Step S33 comprises the following steps: Step S331: querying the plant component space distribution prediction map according to the plant component database, and matching a solvent to obtain plant optimal solvent type data; Step S332: performing temperature sensitivity analysis according to the flavonoid-saponin content prediction value, and constructing an extraction curve of 40℃ constant temperature extraction for 120 minutes to obtain a plant extraction temperature-decay curve; Step S333: performing texture analysis on the tissue hardness characteristics of the real-time plant collection image to obtain a mechanical crushing strength parameter; Step S334: performing small-batch extraction experiments on the plant optimal solvent type data, the site optimization instruction, the plant extraction temperature-decay curve and the mechanical crushing strength parameter to obtain plant predicted yield data.

7. The method of claim 5, wherein the image processing-based plant extraction parameter setting method is characterized by, Step S34 comprises the following steps: Step S341: acquiring plant actual yield data; Step S342: performing difference calculation on the plant actual yield data and the plant predicted yield data to obtain a model difference error matrix; Step S343: performing reinforcement learning on the plant predicted yield data, and mapping the model weight to obtain a parameter prediction accuracy compensation; Step S344: Process parameter iteration is performed on the parameter prediction accuracy compensation and the model differential error matrix to obtain optimized process parameter data.

8. The image processing-based plant extraction parameter setting method according to claim 1, wherein, Step S4 includes the following steps: Step S41: The optimized process parameter data is modularly packaged, and an active ingredient enrichment thermodynamic map is constructed; Step S42: The active ingredient enrichment thermodynamic map is superimposed with the best extraction part, and process parameter prompts are performed to obtain a plant extraction process database; Step S43: A decision report is constructed for the plant extraction process database to generate a plant extraction process decision report.

9. The method of claim 8, wherein the method further comprises: Step S41 includes the following steps: Step S411: The optimized process parameter data is iteratively optimized on the plant process parameter prediction model to obtain a plant process optimization prediction model; Step S412: Input data is obtained, wherein the input data includes raw material images, component distribution maps, and yield data; Step S413: The plant process optimization prediction model is modularly packaged based on the input data, and when the input data is original image data, part optimization data is obtained; when the input data is a component distribution map, solvent-temperature-crushing parameters are obtained; and when the input data is yield data, efficacy potential score data is obtained; Step S414: The part optimization data, solvent-temperature-crushing parameters, and efficacy potential score data are three-dimensionally reconstructed to obtain an active ingredient enrichment thermodynamic map.

10. An image processing-based plant extraction parameter setting system, characterized by, A plant extraction parameter setting system based on image processing for performing the plant extraction parameter setting method based on image processing as claimed in claim 1, comprising: A multi-modal feature construction module for deploying multi-modal data acquisition sensors on plant raw materials and performing multi-scale data fusion to construct a plant multi-source feature dataset, specifically: multi-scale and multi-modal data acquisition on plant raw materials is performed by a hyperspectral imager and a high-resolution microscopic imaging system, and a plant multi-source feature dataset is constructed based on a spectral feature vector, a morphological feature vector, and plant component distribution data; A three-dimensional feature modeling and parameter prediction module for constructing a plant three-dimensional feature space matrix based on the plant multi-source feature dataset, specifically: a three-dimensional feature space matrix is obtained based on plant time dimension encoding data, plant spatial topology encoding data, and plant component distribution data; a plant three-dimensional feature space matrix is modeled to obtain a plant process parameter prediction model; A real-time image prediction and reinforcement learning optimization module for acquiring real-time plant acquisition images; the real-time plant acquisition images are input into the plant process parameter prediction model for component prediction to obtain a plant component spatial distribution prediction map; the plant component spatial distribution prediction map is verified by small-batch extraction experiments to obtain plant predicted yield data; reinforcement learning is performed on the plant predicted yield data, and model weights are relatedly mapped to obtain parameter prediction accuracy compensation; process parameter iteration is performed on the parameter prediction accuracy compensation and the plant predicted yield data to obtain optimized process parameter data; The process parameter packaging and decision visualization module is used for iteratively training a plant process parameter prediction model with optimized process parameter data, and performing module packaging on the plant process parameter prediction model with input data, performing three-dimensional reconstruction according to output data, and obtaining an active ingredient enrichment heat map; superimposing the active ingredient enrichment heat map on the best extraction part, and performing process parameter prompting, to obtain a plant extraction process decision report.

Images