Method and device for detecting plant diseases and insect pests in real time based on computer vision

Abstract

The invention belongs to the technical field of disease and insect pest recognition and detection, and particularly relates to a plant disease and insect pest real-time detection method and device based on computer vision, and the method comprises the steps: obtaining a plant image, and setting the frequency of an obtaining action according to a user demand; constructing a recognition model based on transfer learning; constructing an image preprocessing module, automatically cutting the obtained plant image into a non-plant region image, and performing multispectral feature fusion on the remaining image; inputting the image after multispectral feature fusion into a recognition model, outputting an image set with pest and disease damage features, extracting the image set, and sending a lesion feature region to a cloud end to reduce flow consumption; and automatically updating a local identification model through an incremental learning mechanism, and carrying out plant life cycle management. According to the method, the edge calculation recognition model is deployed in a low-power-consumption embedded mode, only the lesion feature area is uploaded to the cloud after background interference is removed, and good performance is achieved in the aspects of real-time performance and background interference resistance.

Description

Technical Field

[0001] This invention belongs to the field of pest and disease identification and detection technology, specifically relating to a real-time detection method and device for plant pests and diseases based on computer vision. Background Technology

[0002] Desktop plants are far more than just simple decorations. The green environment they provide helps improve attention and cognitive function, and reduces mental fatigue. For office workers or students who need to concentrate for long periods of time, desktop plants are a natural "productivity tool." However, users of desktop plants often lack professional knowledge about pests and diseases. Nowadays, the organic integration of desktop plants with computer vision has become a new trend, which helps to improve the situation where users of desktop plants are not good at managing plants.

[0003] Applying computer vision technology to the management of pests and diseases in desktop potted plants is a technical approach that allows machines to replace human eyes for observation and diagnosis, and provides data support for further precise processing, such as collecting images of plants on desktop potted plants. A Convolutional Neural Network (CNN) model is trained using a large dataset of labeled images of pests and diseases (e.g., healthy leaves, leaves infected with aphids, leaves with powdery mildew, etc.). The trained model can analyze new images uploaded by users to determine whether plants are healthy with high accuracy.

[0004] The following areas still need improvement in the process of computer vision: 1. Online vision devices acquire plant images of desktop potted plants, which requires continuous high-definition image transmission to the cloud. This results in high power consumption and high bandwidth costs for the vision devices. In addition, the high latency of cloud processing makes it difficult to meet real-time requirements.

[0005] 2. Online vision devices may capture images of clutter next to the potted plants on the desktop. This clutter may cause background interference, making it difficult to capture subtle lesion features on the leaves of some small potted plants. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time detection method and device for plant diseases and pests based on computer vision. It features a low-power embedded edge computing recognition model that removes background interference and uploads only the lesion feature area to the cloud, demonstrating good performance in terms of real-time performance and resistance to background interference.

[0007] The specific technical solution adopted by this invention is as follows: A real-time detection method for plant diseases and pests based on computer vision includes: Acquire plant images, and set the frequency of this acquisition action according to user needs; Construct a recognition model based on transfer learning; An image preprocessing module is constructed to automatically crop non-plant areas from the acquired plant images, and the remaining images are fused with multispectral features. The image after multispectral feature fusion is input into the recognition model, which outputs an image set with pest and disease characteristics, and then performs... Extract and send the lesion feature area to the cloud to reduce data consumption; The local recognition model is automatically updated through an incremental learning mechanism, and plant life cycle management is performed.

[0008] As an optional approach, the frequency of the acquisition action specifically includes: Daily maintenance involves taking images of the plants at least once a day using computer vision equipment. Alternatively, real-time monitoring can be performed by capturing plant images at time intervals t using computer vision equipment, with illumination provided during image capture to obtain plant images under near-infrared and visible light spectral conditions.

[0009] As an optional approach, the transfer learning-based recognition model includes: Input a plant image and preprocess it to a standard size to normalize the pixel distribution; We adopted the plug-and-play mode of MobileNetV3+ lightweight attention module and mapped the original image into a feature vector containing semantic information through hierarchical convolution operations. Among them, the feature vectors include powdery mildew features, spider mite and aphid features, and anthracnose features; The feature vector is concatenated with Backbone's global features to form the final feature vector used for classification. .

[0010] As an optional solution, the image preprocessing module includes: The image preprocessing module is trained using a pre-labeled plant image dataset to automatically crop non-plant areas from the acquired plant images. In the CNN neural convolutional network, semantic segmentation is selected for data annotation. A standard dataset containing plant images is created, and each pixel is labeled as either a plant or the background to extract regions with high lesion probability. Backpropagation optimizes network weights, which serves to train the CNN neural convolutional network, thus obtaining the image preprocessing module.

[0011] As an optional approach, the multispectral feature fusion specifically includes: After background segmentation, a masking process is performed using an image preprocessing module to obtain a binary mask image. Let the final output be the image obtained after masking. sigmoidBackground pixel confidence map after activation function; The plant mask is obtained and then fed into the base convolutional layer along with the original infrared image.

[0012] As an optional solution, the The extraction specifically includes: Edge computing devices transmit feature maps of pest and disease images via Wi-Fi or Bluetooth. The region is uploaded to the cloud, and the feature map is displayed. The region can contain .

[0013] As an optional approach, the incremental learning mechanism includes: Users process alarm messages one by one, and when they confirm that a plant has been affected by pests or diseases, they spray the corresponding pesticide on the plant. If a false alarm is confirmed, a false alarm feedback message is sent to the edge computing device to automatically update the local recognition model.

[0014] As an optional approach, the plant lifecycle management includes: A plant lifecycle management report is generated for each edge computing device on a daily basis. Record daily false alarm events and alarm information The false alarm rate was obtained. A continuous line graph of the daily false alarm rate was plotted to show the trend of the false alarm rate at different growth stages of the plant.

[0015] An electronic device, the electronic device comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to perform the real-time detection method for plant diseases and pests.

[0016] A real-time plant disease and pest detection device includes: The main body of a smart flowerpot for planting plants; A multispectral camera and an adjustable spectrum LED are fixed to the top of the smart flowerpot body. The multispectral camera is used to acquire plant images in the smart flowerpot body, and the adjustable spectrum LED is used to provide light with a wavelength range of 400nm to 1000nm. A soil moisture sensor and an ambient light sensor are fixed inside the main body of the smart flowerpot; An edge computing module is used to process the plant image. The edge computing module is electrically connected to the multispectral camera, the soil moisture sensor, the ambient light sensor, and the tunable spectral LED. The edge computing module uploads the lesion feature areas obtained after processing the plant image to the cloud or the user terminal.

[0017] The technical effects achieved by this invention are as follows: This invention differs from general image recognition, specifically addressing the challenges of background interference and micro-symptom detection in desktop potted plant scenarios. It also differs from agricultural greenhouse solutions, as it optimizes the model and transmission mechanism for low-computing-power embedded devices. This meets users' needs for low-power embedded deployment, high-precision recognition in complex desktop environments, and early micro-symptom detection of pests and diseases, demonstrating excellent performance in real-time performance and resistance to background interference.

[0018] This invention provides a real-time detection device that operates under multispectral conditions. It can supplement light with a ring path to solve the problem of uneven lighting on the desktop. It identifies the flowerpot boundary through an edge computing sensor, automatically crops the image of non-plant areas, reduces the amount of computation, and uploads only the disease feature area to the cloud after removing background interference, thereby optimizing the real-time performance of pest and disease detection.

[0019] The system provided by this invention offers a three-level response mechanism to handle different projects. It can quickly identify diseased and pest-infested plants at the local edge and also upload the lesion feature areas to the cloud. It flexibly combines edge computing with traditional cloud computing and utilizes an incremental learning mechanism to automatically update the local recognition model when the user confirms a false alarm, thereby improving the accuracy of personalized recognition. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of the intelligent potted plant of the present invention; Figure 2 This is a system block diagram of a real-time plant disease and pest detection device according to Embodiment 1 of the present invention; Figure 3 This is a flowchart of the real-time detection method for plant diseases and pests in Embodiment 2 of the present invention; Figure 4 This is a flowchart of the recognition model based on transfer learning in Embodiment 2 of the present invention; Figure 5 This is a schematic diagram of the electronic device in Embodiment 2 of the present invention; Figure 6 This is a system block diagram of the real-time plant disease and pest detection system in Embodiment 3 of the present invention.

[0021] The attached diagram lists the components represented by each number as follows: 1. Smart flowerpot body; 2. Edge computing module; 3. Multispectral camera; 4. Soil moisture sensor; 5. Ambient light sensor; 6. Adjustable spectrum LED. Detailed Implementation

[0022] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0023] DIY smart plant kits offer tech enthusiasts a way to organically combine leisure and green living. Users can assemble and program the kits themselves, with the aim of enabling even those without experience to successfully cultivate plants and promoting the popularization of green lifestyles.

[0024] by Figure 1 Taking the smart bonsai as an example, the shape of the smart bonsai can be customized by 3D printing technology and then sprayed with the specified appearance graphics, which is both beautiful and practical. The suspended pole above the smart bonsai can be equipped with LED lights, which can provide light to the plant during a specified time period under the control of a preset program.

[0025] Example 1: like Figure 2 As shown, a real-time plant disease and pest detection device includes: Smart flowerpot body 1 for planting plants; A multispectral camera 3 and an adjustable spectrum LED 6 are fixed to the top of the smart flowerpot body 1. The multispectral camera 3 can be an industrial hyperspectral imaging camera of model SPECIMFX10, used to acquire plant images in the smart flowerpot body 1. The adjustable spectrum LED 6 provides two illumination conditions: visible light and near-infrared light, with a light wavelength range of 400nm to 1000nm. The soil moisture sensor 4 and ambient light sensor 5 are fixed inside the main body 1 of the smart flowerpot. The soil moisture sensor 4 can be a USR-SSO10 or TL-SEN202-TH model soil temperature and humidity sensor, equipped with RS485 / MODBUS-RTU interface communication, supporting multi-device networking (such as cascading 32 devices), and the communication distance can reach 2000 meters. It is used to obtain the temperature and humidity signal inside the main body 1 of the smart flowerpot. The ambient light sensor 5 can be a TSL25911 model high-sensitivity digital ambient light sensor, which outputs light intensity data through the I2C interface (fixed address 0x29) to obtain the ambient light intensity signal of the plant. Edge computing module 2 is used for processing plant images. It utilizes the MA35D development board to implement edge computing. Employing a high-density, high-speed circuit board design, it integrates the MA35D1 chip on a 37mm*39mm board. The board includes built-in DDR, eMMC / Nand, E2PROM, and discrete power supply circuits. Equipped with dual-core Cortex-A35 and Cortex-M4 processors, it boasts powerful processing capabilities and rich communication interfaces. Edge computing module 2 identifies models through algorithmic components and connects to a multispectral camera 3, soil moisture sensor 4, ambient light sensor 5, and tunable spectral LED 6 via a data transmission interface. This enables it to receive temperature and humidity signals, ambient light intensity signals, plant images, and adjust light intensity. Edge computing module 2 then uploads the lesion feature areas obtained after processing the plant images to the cloud or the user terminal.

[0026] Example 2: This example uses succulents, pothos, and miniature roses as examples. They are cultivated within a set number of days, given sufficient water and nutrients, and the changes in the leaves of these plants are observed.

[0027] like Figures 3-4 As shown, a real-time detection method for plant diseases and pests based on computer vision includes the following steps: Timed image acquisition: Acquires plant images. The frequency of this acquisition action can be set according to user needs, for example: Daily maintenance involves taking images of the plants at least once a day using computer vision equipment, with the shooting time set at noon to ensure sufficient light for the images to be obtained. During midday shooting, we proactively retrieved national weather data from the official website of the National Meteorological Administration, which could be linked to local weather information. During periods of cloudy weather, we turned on LED lights to provide auxiliary lighting and illumination. Alternatively, real-time monitoring can be performed by capturing plant images at time intervals t using computer vision equipment. During the capture of plant images, illumination is provided by a ring of 6 LED lights in a distributed manner, with the wavelength range of the illumination being 400nm to 1000nm, so as to acquire plant images under near-infrared and visible light spectral conditions respectively. Building a recognition model: A lightweight pest and disease recognition model based on transfer learning is built using a hybrid network of MobileNetV3 and self-attention mechanism. This recognition model has very few parameters and very low computational cost (FLOPs), making it suitable for deployment on smart potted plants for edge computing. This lightweight pest and disease identification model uses publicly available benchmark datasets (such as PlantVillage, AIChallenger 2018) for rapid validation and pre-training of the model architecture. For example, it divides the model into 5 categories: healthy, powdery mildew, spider mites, aphids and anthracnose, and provides at least 3,000 carefully labeled original images for each category as the training set. The plug-and-play mode of MobileNetV3+ lightweight attention module was selected. For example, the Backbone (feature extractor) maps the original image into a feature vector containing semantic information through hierarchical convolution operations. The process of capturing pest and disease features (powdery mildew, spider mites, aphids, anthracnose) is as follows: Original image preprocessing: Assuming the input plant image is The input plant image is preprocessed to a standard size to standardize the pixel distribution, and the following formula (1) is used to calculate it: Formula (1) in, To input the height value of the plant image, Input the width value of the plant image. =3 represents the number of RGB channels. For input plant images at location ( ),aisle ( The pixel value of ) For training centralized channel The mean, For training centralized channel The standard deviation is used to eliminate interference from different light intensities; Hierarchical feature extraction: Backbone extracts low-level (texture, edges) to high-level (lesion shape, color) features step by step through multiple convolutional blocks (such as Conv-BN-ReLU). Taking a typical CNN convolutional operation as an example: S1, Basic Convolutional Layer Operations: Let the first m The input to a convolutional layer is a feature map. ( (Number of channels), calculate the output feature map. Formula (2) is: Formula (2) in, For example, the kernel size. =3x3 , For the first Layer convolution kernel in aisle,( The weight of the position. The bias of the i-th output channel is used to output the feature map. Indicates the location ( ),aisle ( The set of pixel values, For activation function, Used to enhance linear expression; S2, Residual Connection: Using ResNet as an example, this addresses the degradation of deep networks; Assume the first The layer is a residual block, and the output is a feature map. Input features need to be superimposed (Jump connection), the superposition calculation is as follows: (3) Formula (3) in, for connect; If the number of channels changes, adjust using a 1×1 convolution. S3. Quantification of Pest and Disease Specific Characteristics: High-level feature maps output by the Backbone It contains semantic information about diseases and pests, which is recorded as global features. The features need to be quantified according to the visual characteristics of different diseases and pests. (I) Characteristics of powdery mildew, mainly manifested in white powdery substances, marked by high-brightness areas; Powdery mildew appears as a blurry white area on the leaf surface in the image (around 255 in RGB). The quantitative formula for powdery mildew is... The following formula (4) is used for calculation: Formula (4) in, The region in the feature map is suspected to be powdery mildew (high-brightness areas are filtered by threshold). for The formula for the number of inner pixels quantifies the white powder characteristics by calculating the consistency of RGB channel values ​​(the smaller the difference, the purer the white). (II) Characteristics of spider mites and aphids: mainly small insects, identified by color and texture; The core characteristic of spider mites (small red / brown dots) and aphids (small green / black dots) is "small-scale, high-contrast areas." The quantification formula for spider mites is... The formula for quantifying aphids is: The calculations for both are as follows: (5) and (6): Formula (5) Formula (6) in, (∙) represents the feature map The variance of the region is large when the edge texture of the insect is clear, while the variance of the red spider depends on... Channel mean (high red component), aphid dependence Channel mean (high green component); (III) Anthracnose characteristics: mainly manifested as brown necrotic spots on the leaves, with clear edges of these brown necrotic spots; Anthracnose manifests as brown, circular spots on leaves (with a large gradient at the edges), and the quantitative formula is: The following formula (7) is used for calculation: Formula (7) in, for Orientation gradient, for Orientation gradient, The browning intensity is quantified by the area of ​​the spots and the mean channel value, especially in anthrax. Channel values ​​are higher than those of healthy leaves; S4. Final feature vector concatenation: The above-mentioned pest-specific features are concatenated with Backbone's global features to form the final feature vector used for classification. The following formula (8) is used for calculation: Formula (8) in, For features obtained after global pooling of the Backbone, such as as the output of GAP, Used in subsequent pest and disease classifiers, it serves as a lightweight model that can be embedded in edge computing devices at a low cost. Target image preprocessing: An image preprocessing module is constructed based on a CNN neural convolutional network. The image preprocessing module is trained with a pre-labeled plant image dataset to automatically crop non-plant areas from the acquired plant images. The purpose is to clean the acquired raw image data and focus on the main lesion areas. (IV) Background segmentation: In the CNN neural convolutional network, the semantic segmentation method is selected. First, data annotation is performed to create a standard dataset containing plant images and label each pixel as either a plant or the background. High-probability lesion regions such as leaf tips and leaf edges are extracted first. For example, on datasets such as Pascal VOC or Cityscapes, add specific category annotations about flower pots and plants, treating flower pots as a "non-plant category target". This allows semantic segmentation algorithms to distinguish them and remove unwanted parts. In semantic segmentation, labeled images are used to optimize network weights through backpropagation, effectively training a CNN (Convolutional Neural Network) to obtain the image preprocessing module. Network C At this point, it can be Network C Preprocessing for raw plant images; Multispectral feature fusion: This embodiment already has a relatively lightweight pest and disease identification model and image preprocessing module. However, the original images were taken under at least two spectral conditions, which resulted in two types of original images: one is a common light image and the other is an infrared image. By comparing multispectral features, the pest and disease characteristics of the same leaf can be compared, reducing errors. However, pest and disease identification models can generally only recognize common light images, so infrared images cannot be directly input into the pest and disease identification model. Here, the infrared images need to be masked. Masking is a process that uses an image preprocessing module to obtain a binary mask image after background segmentation. Binary Mask Assuming its final output is after sigmoid The background pixel confidence map after activation function is as follows BPM_vis [ The following formula (9) is used for calculation: Formula (9) in, This mask is used to indicate which areas can be excluded from the image after it has been processed by the image preprocessing module. Obtain plant cover Plant Mask This can be done in the image preprocessing module. Network C The result is obtained directly from the prediction, and the calculation process is as follows: Formula (10): Formula (10) If it is a binary image with a background probability or background confidence greater than 0.5, then using it to exclude background clutter is very intuitive; Lightweight model inference: To optimize real-time performance, this... The plant mask is applied to the corresponding infrared image, and the obtained plant mask (existing in the form of a feature map) and the original infrared image are input together into the basic convolutional layer operation. That is, the image after multispectral feature fusion is input into the recognition model, and the output is an image set with pest-specific features. O and the image set O They are divided into two categories: common pests and diseases and suspected pests and diseases, and the procedures for each are as follows: (V) Device alarm: For common pest and disease images obtained by local edge computing, if the confidence level of the common pest and disease images is greater than 95%, the edge computing device can send alarm information to the user's mobile phone via low power WIFI or Bluetooth, or it can also use a built-in buzzer to emit a piercing sound for alarm purposes. (VI) Extraction: For suspected pest and disease images obtained by the identification model on the edge computing device, where the confidence level of these suspected pest and disease images is greater than 70% and less than 95%, the edge computing device can extract the feature maps of the suspected pest and disease images via low-power Wi-Fi or Bluetooth. The region is uploaded to the cloud, and the feature map is displayed. The region can contain This enables data compression and transmission, which can reduce traffic by 90%, helping to reduce traffic costs, overcome the high latency problem of traditional cloud processing, and actively meet real-time requirements. Incremental learning mechanism: Users process the alarm information they receive one by one, and spray the corresponding pesticide on the plant when they confirm that the plant has been affected by pests or diseases; If a false alarm is confirmed, a false alarm feedback message is sent to the edge computing device to automatically update the local recognition model, delete the convolution kernel corresponding to the pest feature, and improve the accuracy of personalized recognition. Lifecycle Management: A plant lifecycle management report is generated for each edge computing device on a daily basis, recording false alarm events each day. and alarm information The false alarm rate was obtained. The system also plots a continuous line graph of the daily false alarm rate, allowing users to visually see the trend of the false alarm rate at different growth stages of the plant. This enables them to increase the frequency of management during periods of high false alarm rates, making the identification model in this embodiment more closely aligned with the changes in the plant life cycle.

[0028] An electronic device comprising: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores a computer program that can be executed by at least one processor, which enables the at least one processor to perform a real-time detection method for plant diseases and pests.

[0029] The processors in the aforementioned electronic devices can be of various types, such as CPUs, GPUs, or TPUs, to adapt to different computing needs and ensure efficient processing of multidimensional disaster data. The memory can be of various types, such as RAM, ROM, or SSDs, to store large amounts of data, support fast read and write operations, and ensure stable system operation. In addition, the electronic devices are equipped with high-precision sensors to monitor environmental changes in real time, ensuring the accuracy and timeliness of data acquisition, as well as arithmetic units, input devices, and output devices. The arithmetic unit can be an FPGA or ASIC, responsible for high-speed parallel computing. Input devices such as keyboards and touch screens facilitate operation, while output devices such as displays and printers intuitively display results. The entire system works in tandem to improve the response speed and decision support capabilities of the real-time plant disease and pest detection device.

[0030] Example 3: like Figure 6 As shown, a real-time plant disease and pest detection system includes: The image capture module captures plant images in real time. The supplemental lighting module features a ring-shaped structure with six LED lights (including near-infrared light sources) surrounding the plant to illuminate it along the ring path. The edge computing module, through the real-time detection method for plant diseases and pests in the embedded implementation of the algorithm, provides a three-level response mechanism. After inputting a plant image, it can output common disease and pest characteristics and suspected disease and pest characteristics on the local edge computing side, and upload the lesion feature area to the cloud. The cloud-based expert system receives the characteristic areas of the lesions and processes them asynchronously to identify rare diseases or new plant species.

[0031] The processing results of the three-level response mechanism are shown in Table 1 below.

[0032] Table 1. Results of the Level 3 Response Mechanism By comparing the data in Table 1, it can be seen that the local edge computing speed for processing plant images is fast, and the network latency of uploading data to the cloud can be almost ignored. It can provide the output results of pest and disease characteristics within 2 seconds, which actively helps users manage plants.

[0033] This system was compared with a traditional cloud solution. The same 1,000 original plant images were input into both systems, and the various indicators of the two systems were statistically analyzed. The results are shown in Table 2 below.

[0034] Table 2 Statistical Results The data above shows that, unlike general image recognition, it specifically addresses the challenges of background interference and micro-symptom detection in desktop potted plant scenarios. It also differs from agricultural greenhouse solutions, as it optimizes the model and transmission mechanism for low-computing-power embedded devices. This meets users' needs for low-power embedded deployment, high-precision recognition in complex desktop environments, and early micro-symptom detection of pests and diseases, demonstrating excellent performance in real-time performance and resistance to background interference.

[0035] The above description is merely an optional embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention, unless otherwise specified or limited, shall be implemented according to conventional means in the art.

Claims

1. A real-time detection method for plant diseases and pests based on computer vision, characterized in that, include: Acquire plant images, and set the frequency of this acquisition action according to user needs; Construct a recognition model based on transfer learning; An image preprocessing module is constructed to automatically crop non-plant areas from the acquired plant images, and the remaining images are fused with multispectral features. The image after multispectral feature fusion is input into the recognition model, which outputs an image set with pest and disease characteristics, and then performs... Extract and send the lesion feature area to the cloud to reduce data consumption; The local recognition model is automatically updated through an incremental learning mechanism, and plant life cycle management is performed.

2. The method for real-time detection of plant diseases and pests according to claim 1, characterized in that, The frequency of the acquisition action specifically includes: Daily maintenance involves taking images of the plants at least once a day using computer vision equipment. Alternatively, real-time monitoring can be performed by capturing plant images at time intervals t using computer vision equipment, with illumination provided during image capture to obtain plant images under near-infrared and visible light spectral conditions.

3. The method for real-time detection of plant diseases and pests according to claim 1, characterized in that, The recognition model based on transfer learning includes: Input a plant image and preprocess it to a standard size to normalize the pixel distribution; We adopted the plug-and-play mode of MobileNetV3+ lightweight attention module and mapped the original image into a feature vector containing semantic information through hierarchical convolution operations. Among them, the feature vectors include powdery mildew features, spider mite and aphid features, and anthracnose features; The feature vector is concatenated with Backbone's global features to form the final feature vector used for classification. .

4. The method for real-time detection of plant diseases and pests according to claim 1, characterized in that, The image preprocessing module includes: The image preprocessing module is trained using a pre-labeled plant image dataset to automatically crop non-plant areas from the acquired plant images. In the CNN neural convolutional network, semantic segmentation is selected for data annotation. A standard dataset containing plant images is created, and each pixel is labeled as either a plant or the background to extract regions with high lesion probability. Backpropagation optimizes network weights, which serves to train the CNN neural convolutional network, thus obtaining the image preprocessing module.

5. The method for real-time detection of plant diseases and pests according to claim 1, characterized in that, The multispectral feature fusion specifically includes: After background segmentation, a masking process is performed using an image preprocessing module to obtain a binary mask image. Let the final output be the image obtained after masking. sigmoid Background pixel confidence map after activation function; The plant mask is obtained and then fed into the base convolutional layer along with the original infrared image.

6. The method for real-time detection of plant diseases and pests according to claim 1, characterized in that, The The extraction specifically includes: Edge computing devices transmit feature maps of pest and disease images via Wi-Fi or Bluetooth. The region is uploaded to the cloud, and the feature map is displayed. The region can contain .

7. The method for real-time detection of plant diseases and pests according to claim 1, characterized in that, The incremental learning mechanism includes: Users process alarm messages one by one, and when they confirm that a plant has been affected by pests or diseases, they spray the corresponding pesticide on the plant. If a false alarm is confirmed, a false alarm feedback message is sent to the edge computing device to automatically update the local recognition model.

8. The method for real-time detection of plant diseases and pests according to claim 1, characterized in that, The plant life cycle management includes: A plant lifecycle management report is generated for each edge computing device on a daily basis. Record daily false alarm events and alarm information The false alarm rate was obtained. A continuous line graph of the daily false alarm rate was plotted to show the trend of the false alarm rate at different growth stages of the plant.

9. A real-time plant disease and pest detection system, employing the real-time plant disease and pest detection method according to any one of claims 1-8, characterized in that, include: The image capture module captures plant images in real time. The supplemental lighting module uses LED lights to illuminate the plants along a circular path; The edge computing module, through the embedded algorithm of the real-time detection method for plant diseases and pests, provides a three-level response mechanism. After inputting a plant image, it outputs common disease and pest features and suspected disease and pest features on the local edge computing side, and uploads the disease and pest feature area to the cloud. The cloud-based expert system receives information about pest and disease characteristics in a region and processes it asynchronously.

10. A real-time plant disease and pest detection device, using the real-time plant disease and pest detection method according to any one of claims 1-8, characterized in that, include: Smart flowerpot body for planting plants (1); A multispectral camera (3) and an adjustable spectrum LED (6) are fixed to the top of the smart flowerpot body (1). The multispectral camera (3) is used to acquire plant images in the smart flowerpot body (1), and the adjustable spectrum LED (6) is used to provide light with a wavelength range of 400nm to 1000nm. A soil moisture sensor (4) and an ambient light sensor (5) are fixed inside the main body (1) of the smart flowerpot. An edge computing module (2) is installed on the main body (1) of the smart flowerpot. The edge computing module (2) is electrically connected to the multispectral camera (3), the soil moisture sensor (4), the ambient light sensor (5), and the adjustable spectrum LED (6).

Images