Plant scanning method and plant scanning AI glasses
By constructing a local feature vector library and dynamically adjusting the threshold of a smartphone app, the problem of plant identification in areas without base station coverage for AI glasses was solved, achieving high-precision and environmentally adaptive plant identification and improving the continuity and reliability of field surveys.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CRETE YUNKE (GUANGDONG) TECHNOLOGY CO LTD
- Filing Date
- 2026-03-27
- Publication Date
- 2026-06-26
AI Technical Summary
Existing plant identification AI glasses rely on cloud-based comparison, making them unusable in areas without base station coverage. Furthermore, their identification thresholds are fixed and cannot be flexibly adjusted based on operator experience, regional differences, and real-time weather conditions, resulting in insufficient identification accuracy and adaptability.
A local feature vector library is built, and communication is established between a smartphone APP and AI glasses to collect images in real time and perform target detection and feature matching. Combined with manual threshold adjustment, standard reference value switching and weather adaptive module, the confidence decision threshold is dynamically adjusted to achieve accurate identification and positioning of plant targets.
Achieving high species fine-grained identification accuracy and adaptability to all environmental elements without network connection improves the continuity and reliability of long-term, cross-regional plant surveys in the field.
Smart Images
Figure CN122293783A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of plant scanning AI glasses technology, specifically relating to a plant scanning method and plant scanning AI glasses. Background Technology
[0002] Cordyceps sinensis is a precious medicinal fungal complex distributed in the alpine meadows of the Qinghai-Tibet Plateau. It has extremely high economic and scientific research value. However, its stroma usually only protrudes 2-5 centimeters above the ground, and its color is very similar to that of withered grass and soil. Moreover, it grows in remote, uninhabited areas at an altitude of over 4,000 meters. Traditional search methods rely entirely on collectors to bend over and crawl, and identify it with the naked eye. A single person can search less than 0.3 hectares per day, and the missed detection rate is as high as 70%.
[0003] In recent years, AI glasses technology integrating artificial intelligence and augmented reality has been gradually introduced into this field, realizing a real-time recognition experience where what you see is what you know. However, existing plant recognition AI glasses are heavily reliant on networks, and the recognition scheme requires uploading images to the cloud for comparison with the database. However, many cordyceps production areas are located in areas without base station coverage, making it impossible to establish a stable connection. Moreover, core parameters such as recognition thresholds are fixed after leaving the factory and cannot be flexibly adjusted according to the operator's experience, regional differences, and real-time weather conditions, which limits the adaptability of the equipment in different scenarios.
[0004] To address the aforementioned issues, this application presents a plant scanning method and plant scanning AI glasses. Summary of the Invention
[0005] To address the shortcomings of the prior art mentioned in the background section, this application proposes a plant scanning method and plant scanning AI glasses. First, a local feature vector library is constructed and connected to the AI glasses via a smartphone app. Then, a threshold adjustment command is generated via the smartphone app and sent to the AI glasses to update the current confidence threshold θ. Next, the AI glasses acquire a real-time sequence of visual field images, performing target detection and feature matching frame by frame. Feature vectors of candidate regions are extracted to calculate the confidence C of the target species. Then, the calculated confidence C is compared with the currently effective confidence threshold θ. If C > θ, the target is identified, and an augmented reality bounding box and directional guidance are generated to solve the problems in the background section.
[0006] To achieve the above objectives, this application provides a plant scanning AI glasses, including an AI glasses terminal and a smartphone APP. The AI glasses terminal integrates a side-end real-time visual acquisition and processing module, an optical display module, and an environment adaptive perception unit for collecting environmental parameters of the work site. The side-end real-time visual acquisition and processing module integrates a 4K global shutter camera and an edge AI computing unit. The edge AI computing unit deploys a lightweight deep learning model and stores a local feature vector library, enabling real-time plant target recognition and localization without a network connection.
[0007] The optical display module is used to accurately anchor and superimpose the recognition result in the form of a dynamic marker box onto the corresponding spatial position of the target object in the wearer's field of vision;
[0008] The smartphone app establishes a two-way data link with the AI glasses via Bluetooth Low Energy to support parameter delivery and status upload. The smartphone app also has a built-in threshold manual adjustment module, a standard reference value switching module, and a weather recognition and threshold adaptive module.
[0009] In a preferred embodiment based on the above scheme, the threshold manual adjustment module is used to receive user input and generate a confidence judgment threshold adjustment instruction, which is then sent to the AI glasses to update the threshold parameters.
[0010] The standard reference value switching module is used to switch the feature vector library loaded on the AI glasses and the corresponding threshold preset scheme.
[0011] The weather recognition and threshold adaptation module is used to acquire weather data of the work area and environmental parameters returned by the AI glasses, dynamically generate threshold correction values based on the adaptive algorithm, and send them to the AI glasses.
[0012] Based on the above scheme, the lightweight deep learning model built into the edge AI computing unit is compressed using knowledge distillation and INT8 quantization technology, with a model parameter size of less than 8MB and an end-to-end inference latency of less than 150ms per frame.
[0013] Based on the above scheme, the preferred embodiment of the optical display module adopts an arrayed waveguide scheme with an eye brightness of over 1800 nits. The display marker frame is scaled in real time according to the binocular depth estimation results and maintains visual anchoring with the target object in three-dimensional space.
[0014] A plant scanning method includes the following steps:
[0015] S1: Establish a communication connection between the smartphone APP and the AI glasses, and initialize the local feature vector library and threshold parameters;
[0016] S2: Generate a threshold adjustment command via a smartphone app and send it to the AI glasses to update the current confidence threshold θ;
[0017] S3: Real-time acquisition of visual field image sequences through AI glasses, execution of target detection and feature matching frame by frame, extraction of candidate region feature vectors, and calculation of the confidence C of each candidate region belonging to the target species based on the preset confidence calculation formula;
[0018] S4: Compare the calculated confidence level C with the currently effective confidence level decision threshold θ. If C > θ, it is determined to be a target and an augmented reality marker box and directional guidance are generated and superimposed on the wearer's field of vision through the optical display module.
[0019] S5: Record the geographic coordinates, altitude, discovery time, and image thumbnail of the target plant, generate a local resource distribution record, and export it synchronously via a smartphone APP.
[0020] In a preferred embodiment based on the above scheme, the step of generating a threshold adjustment command via a smartphone app and sending it to the AI glasses to update the current confidence judgment threshold θ includes the following steps:
[0021] S21: The mode is switched through the standard reference value switching module built into the smartphone APP. When switching modes, the smartphone APP sends the corresponding feature vector library and threshold baseline. The AI glasses load the corresponding feature vector library and automatically complete the setting of the confidence judgment threshold to adjust the mode switching.
[0022] S22: Based on actual usage, operate through the threshold setting interface of the smartphone APP, send the confidence target threshold code to the AI glasses, and complete manual fine-tuning;
[0023] S23: Periodically obtain weather forecast data for the work area and on-site lighting, temperature, and humidity parameters transmitted from the AI glasses via a smartphone APP. Calculate the dynamic confidence threshold using the pre-set adaptive algorithm within the weather recognition and threshold adaptation module, and send the calculated dynamic confidence threshold to the AI glasses for weather adaptive adjustment.
[0024] Based on the above scheme, the preferred formula for calculating the confidence level in step S3 is: Where f is the candidate region feature vector, and p is the reference feature vector of the target category in the local feature library under the current activation mode. Let be the cosine similarity function, τ be the offset parameter, and λ be the temperature coefficient. This is the Sigmoid function.
[0025] Based on the above scheme, the target detection in step S3 adopts a two-stage cascaded architecture. The first stage uses spectral prior to quickly screen out candidate regions. The second stage performs fine classification and target box regression within the screened candidate regions and outputs the confidence C of the target species.
[0026] Based on the above scheme, the preferred method for constructing the local feature vector library is as follows: first, collect multi-view and multi-phenological image data of the target species in its native habitat; then, use generative adversarial networks or diffusion models to perform data augmentation and background broadening; and finally, use contrastive learning to pre-train a feature extraction network and generate a feature index.
[0027] Compared with the prior art, the beneficial effects of the present invention are:
[0028] This invention, by constructing a local feature vector library, enables accurate identification in offline environments, making it suitable for use in areas without base station coverage. Through a smartphone app with a manual threshold adjustment module, a standard reference value switching module, and a weather recognition and threshold adaptation module, users can switch modes based on the usage scenario, load the corresponding feature vector library, and automatically set the confidence threshold. Furthermore, users can manually fine-tune the threshold setting via the smartphone app's threshold setting interface, sending the target confidence threshold to the AI glasses. During use, the smartphone app periodically retrieves weather forecast data for the work area and on-site light, temperature, and humidity parameters from the AI glasses. The weather recognition and threshold adaptation module uses a pre-built adaptive algorithm to calculate the dynamic confidence threshold, which is then sent to the AI glasses for weather-adaptive adjustments. The confidence threshold value can be flexibly adjusted based on operator experience, regional differences, and real-time weather conditions. This achieves both high species fine-grained identification accuracy and full environmental element adaptation capability on an extremely lightweight edge device, significantly improving the continuity and reliability of long-term, cross-regional plant surveys in the field. Attached Figure Description
[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0030] Figure 1 This is a schematic diagram of the overall process of a plant scanning method according to the present invention;
[0031] Figure 2 This is a flowchart illustrating step S2 in a plant scanning method according to the present invention. Detailed Implementation
[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0033] Example:
[0034] To address the technical problems raised in the background art, this application provides a preferred embodiment: such as Figures 1-2 As shown, a plant scanning AI glasses includes an AI glasses terminal and a smartphone APP. The AI glasses terminal integrates a side-end real-time visual acquisition and processing module, an optical display module, and an environmental adaptive perception unit for collecting environmental parameters of the work site. The side-end real-time visual acquisition and processing module integrates a 4K global shutter camera and an edge AI computing unit. The edge AI computing unit deploys a lightweight deep learning model and stores a local feature vector library, which can complete plant target recognition and positioning in real time without network connection.
[0035] The optical display module is used to precisely anchor and overlay the recognition results as dynamic marker boxes onto the corresponding spatial position of the target object in the wearer's field of vision;
[0036] The smartphone app establishes a two-way data link with the AI glasses via Bluetooth Low Energy to support parameter delivery and status upload. The smartphone app also has a built-in module for manual threshold adjustment, a module for switching standard reference values, and a module for weather recognition and threshold adaptation.
[0037] The advantages of the above solution are as follows: by building a local feature vector library, accurate identification can be performed in a network-free environment, making it suitable for use in areas without base station coverage. By setting a manual threshold adjustment module, a standard reference value switching module, and a weather recognition and threshold adaptation module in the smartphone APP, the confidence judgment threshold value can be flexibly adjusted according to the operator's experience, regional differences, and real-time weather conditions. High species fine-grained identification accuracy and full environmental element adaptation capability are achieved simultaneously on an extremely lightweight edge device, significantly improving the continuity and reliability of long-term, cross-regional plant survey operations in the field.
[0038] In an optional embodiment, the threshold manual adjustment module is used to receive user input and generate a confidence judgment threshold adjustment instruction, which is then sent to the AI glasses to update the threshold parameters.
[0039] The standard reference value switching module is used to switch the feature vector library loaded on the AI glasses and the corresponding threshold preset scheme;
[0040] The weather recognition and threshold adaptation module is used to acquire weather data of the work area and environmental parameters returned by the AI glasses. It dynamically generates threshold correction values based on the adaptive algorithm and sends them to the AI glasses.
[0041] In an optional embodiment, the lightweight deep learning model built into the edge AI computing unit is compressed using knowledge distillation and INT8 quantization techniques, resulting in a model parameter size of less than 8MB and an end-to-end inference latency of less than 150ms per frame.
[0042] In an optional embodiment, the optical display module employs an arrayed waveguide scheme with an eye brightness exceeding 1800 nits. The display marker frame scales in real time based on binocular depth estimation results and maintains visual anchoring to the target object in three-dimensional space.
[0043] A plant scanning method includes the following steps:
[0044] S1: Establish a communication connection between the smartphone APP and the AI glasses, and initialize the local feature vector library and threshold parameters;
[0045] S2: Generate a threshold adjustment command via a smartphone app and send it to the AI glasses to update the current confidence threshold θ;
[0046] S3: Real-time acquisition of visual field image sequences through AI glasses, execution of target detection and feature matching frame by frame, extraction of candidate region feature vectors, and calculation of the confidence C of each candidate region belonging to the target species based on the preset confidence calculation formula;
[0047] S4: Compare the calculated confidence level C with the currently effective confidence level decision threshold θ. If C > θ, it is determined to be a target and an augmented reality marker box and directional guidance are generated and superimposed on the wearer's field of vision through the optical display module.
[0048] S5: Record the geographic coordinates, altitude, discovery time, and image thumbnail of the target plant, generate a local resource distribution record, and export it synchronously via a smartphone APP.
[0049] The advantages of the above scheme are as follows: it can perform target detection and feature matching frame by frame on the acquired visual field image sequence in a network-free environment, extract the feature vector of the candidate region, and calculate the confidence C of each candidate region belonging to the target species based on the preset confidence calculation formula. Then, the calculated confidence C is compared with the currently effective confidence decision threshold θ. If C > θ, it is determined to be a target and an augmented reality marker box and directional guidance are generated. By superimposing it on the wearer's field of vision through the optical display module, the location of Cordyceps can be quickly identified and marked, which can effectively improve the harvesting efficiency of Cordyceps.
[0050] Furthermore:
[0051] In an optional embodiment, a threshold adjustment instruction is generated via a smartphone app and sent to the AI glasses to update the current confidence decision threshold θ, including the following steps:
[0052] S21: The mode is switched through the standard reference value switching module built into the smartphone APP. When switching modes, the smartphone APP sends the corresponding feature vector library and threshold baseline. The AI glasses load the corresponding feature vector library and automatically complete the setting of the confidence judgment threshold to adjust the mode switching.
[0053] S22: Based on actual usage, operate through the threshold setting interface of the smartphone APP, send the confidence target threshold code to the AI glasses, and complete manual fine-tuning;
[0054] S23: Periodically obtain weather forecast data for the work area and on-site lighting, temperature, and humidity parameters transmitted from the AI glasses via a smartphone APP. Calculate the dynamic confidence threshold using the pre-set adaptive algorithm within the weather recognition and threshold adaptation module, and send the calculated dynamic confidence threshold to the AI glasses for weather adaptive adjustment.
[0055] The advantages of the above solution are as follows: During use, it not only allows for mode switching based on the application scenario, loading the corresponding feature vector library, and automatically setting the confidence threshold, but also enables manual fine-tuning through the threshold setting interface of a smartphone app. The target confidence threshold is encoded and sent to the AI glasses for manual adjustment. Furthermore, during use, the smartphone app periodically obtains weather forecast data for the work area and on-site light, temperature, and humidity parameters transmitted from the AI glasses. The dynamic confidence threshold is calculated using a pre-built adaptive algorithm within the weather recognition and threshold adaptation module, and then sent to the AI glasses for weather-adaptive adjustments. The confidence threshold value can be flexibly adjusted based on operator experience, regional differences, and real-time weather conditions. This achieves both high species fine-grained recognition accuracy and full environmental element adaptability on an extremely lightweight edge device, significantly improving the continuity and reliability of long-term, cross-regional plant surveys in the field.
[0056] In an optional embodiment, the confidence level calculation formula in step S3 is: Where f is the candidate region feature vector, and p is the reference feature vector of the target category in the local feature library under the current activation mode. Let be the cosine similarity function, τ be the offset parameter, and λ be the temperature coefficient. This is the Sigmoid function.
[0057] It should be noted that the core logic of the confidence score calculation formula is to measure the feature matching degree between the candidate region and the target species using cosine similarity, and then convert the similarity into an interpretable confidence score using the sigmoid function. By converting the abstract feature similarity into an intuitive confidence score value, it is easier for users and the system to understand the reliability of the recognition results. The similarity between candidate region features and target reference features is measured, with values ranging from (-1, 1). The offset parameter τ is used to adjust the baseline of similarity, and the temperature coefficient λ is used to control the steepness of the Sigmoid function, thereby adjusting the sensitivity of the confidence score to changes in similarity. Simultaneously, the Sigmoid function... Mapped to the (0,1) interval, the output confidence level C can be directly compared with the confidence decision threshold θ, which meets the engineering requirements of probabilistic decision-making.
[0058] Furthermore:
[0059] In an optional embodiment, the target detection in step S3 adopts a two-stage cascaded architecture. The first stage uses spectral priors to quickly screen out candidate regions. The second stage performs fine classification and target box regression within the screened candidate regions and outputs the confidence C of the target species.
[0060] It should be noted that by using coarse screening and fine detection to achieve an unbalanced allocation of computing resources, and through engineering methods, the complex biometric task is broken down into spectral-level prediction and semantic-level fine detection. This not only utilizes the speed of traditional optical knowledge but also leverages the accuracy of deep learning, perfectly matching the core technical requirements of lightweight, low latency, and offline availability in the solution.
[0061] In an optional embodiment, the local feature vector library is constructed by first collecting multi-view, multi-phenological image data of the target species in its native habitat, then using generative adversarial networks or diffusion models for data augmentation and background broadening, and finally using contrastive learning to pre-train a feature extraction network and generate a feature index.
[0062] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A plant scanning AI glasses, comprising an AI glasses end and a smart phone APP, characterized in that, The AI glasses are equipped with a side-end real-time visual acquisition and processing module, an optical display module, and an environment adaptive perception unit for collecting environmental parameters of the work site. The side-end real-time visual acquisition and processing module integrates a 4K global shutter camera and an edge AI computing unit. The edge AI computing unit is equipped with a lightweight deep learning model and stores a local feature vector library, which can complete plant target identification and positioning in real time without network connection. The optical display module is used to accurately anchor and superimpose the recognition result in the form of a dynamic marker box onto the corresponding spatial position of the target object in the wearer's field of vision; The smartphone app establishes a two-way data link with the AI glasses via Bluetooth Low Energy to support parameter delivery and status upload. The smartphone app also has a built-in threshold manual adjustment module, a standard reference value switching module, and a weather recognition and threshold adaptive module.
2. The plant scanning AI glasses of claim 1, wherein: The threshold manual adjustment module is used to receive user input and generate a confidence judgment threshold adjustment instruction, which is then sent to the AI glasses to update the threshold parameters. The standard reference value switching module is used to switch the feature vector library loaded on the AI glasses and the corresponding threshold preset scheme. The weather recognition and threshold adaptation module is used to obtain weather data of the work area and environmental parameters returned by the AI glasses, dynamically generate threshold correction values based on the adaptive algorithm, and send them to the AI glasses.
3. The plant scanning AI glasses of claim 1, wherein: The lightweight deep learning model built into the edge AI computing unit uses knowledge distillation and INT8 quantization technology for compression, with a model parameter size of less than 8MB and an end-to-end inference latency of less than 150ms per frame.
4. The plant scanning AI glasses of claim 1, wherein: The optical display module adopts an arrayed waveguide scheme, with an eye brightness of over 1800 nits. The display marker frame is scaled in real time based on the binocular depth estimation results and maintains visual anchoring to the target object in three-dimensional space.
5. A plant scanning method, applied to the plant scanning AI glasses according to any one of claims 1-4, characterized in that, Includes the following steps: S1: Establish a communication connection between the smartphone APP and the AI glasses, and initialize the local feature vector library and threshold parameters; S2: Generate a threshold adjustment command via a smartphone app and send it to the AI glasses to update the current confidence threshold θ; S3: Real-time acquisition of visual field image sequences through AI glasses, execution of target detection and feature matching frame by frame, extraction of candidate region feature vectors, and calculation of the confidence C of each candidate region belonging to the target species based on the preset confidence calculation formula; S4: Compare the calculated confidence level C with the currently effective confidence level decision threshold θ. If C > θ, it is determined to be a target and an augmented reality marker box and directional guidance are generated and superimposed on the wearer's field of vision through the optical display module. S5: Record the geographic coordinates, altitude, discovery time, and image thumbnail of the target plant, generate a local resource distribution record, and export it synchronously via a smartphone APP.
6. A plant scanning method according to claim 5, characterized in that: The step of generating a threshold adjustment command via a smartphone app and sending it to the AI glasses to update the current confidence judgment threshold θ includes the following steps: S21: The mode is switched through the standard reference value switching module built into the smartphone APP. When switching modes, the smartphone APP sends the corresponding feature vector library and threshold baseline. The AI glasses load the corresponding feature vector library and automatically complete the setting of the confidence judgment threshold to adjust the mode switching. S22: Based on actual usage, operate through the threshold setting interface of the smartphone APP, send the confidence target threshold code to the AI glasses, and complete manual fine-tuning; S23: Periodically obtain weather forecast data for the work area and on-site lighting, temperature, and humidity parameters transmitted from the AI glasses via a smartphone APP. Calculate the dynamic confidence threshold using the pre-set adaptive algorithm within the weather recognition and threshold adaptation module, and send the calculated dynamic confidence threshold to the AI glasses for weather adaptive adjustment.
7. The method of claim 5, wherein: The confidence level calculation formula in step S3 is as follows: Where f is the candidate region feature vector, and p is the reference feature vector of the target category in the local feature library under the current activation mode. Let be the cosine similarity function, τ be the offset parameter, and λ be the temperature coefficient. This is the Sigmoid function.
8. The method of claim 5, wherein: The target detection in step S3 adopts a two-stage cascaded architecture. The first stage uses spectral priors to quickly screen out candidate regions. The second stage performs fine classification and target box regression within the screened candidate regions and outputs the confidence C of the target species.
9. The method of claim 6, wherein: The method for constructing the local feature vector library is as follows: first, collect multi-view and multi-phenological image data of the target species in its native habitat; then, use generative adversarial networks or diffusion models to perform data augmentation and background broadening; and finally, use contrastive learning to pre-train a feature extraction network and generate a feature index.