Cigar tobacco leaf multi-disease identification and intelligent prevention and control system based on image detection

Abstract

The invention discloses a cigar tobacco multi-disease identification and intelligent prevention and control system based on image detection, which comprises an image acquisition module, a preprocessing module, a multi-disease identification model module and a decision-making and prevention and control module, and all the modules realize real-time communication through a data transmission protocol to form a closed-loop system from disease detection to precise prevention and control. According to the method, high-precision real-time identification of multiple diseases of the cigar tobacco leaves is realized, multiple diseases such as blue mold, black shank, brown spot and bacterial wilt can be synchronously detected, and the problems of misjudgment and missing detection of complex symptoms in a traditional method are solved.

Description

Technical Field

[0001] This invention belongs to the field of agricultural disease identification and control technology, specifically a multi-disease identification and intelligent control system for cigar tobacco leaves based on image detection. Background Technology

[0002] Cigar tobacco leaves are susceptible to a variety of diseases with similar symptoms. Traditional manual detection is inefficient and has a high rate of misjudgment. At the same time, existing image recognition systems are mostly designed for single diseases, which is not comprehensive. They cannot detect multiple diseases such as black shank, red spot, and bacterial wilt at the same time. Furthermore, they are not adaptable to complex backgrounds (such as overlapping leaves and changes in light) and lack a closed-loop system that links disease identification and control, making it difficult to achieve rapid response. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an image detection-based system for the identification and intelligent control of multiple diseases in cigar tobacco leaves, which addresses the shortcomings of the existing technology. This system enables high-precision real-time identification of multiple diseases in cigar tobacco leaves and can simultaneously detect multiple diseases such as blue mold, black shank, red spot disease, and bacterial wilt, overcoming the problems of misjudgment and missed detection of complex symptoms in traditional methods.

[0004] The technical solution of this invention to solve the above technical problems is: an image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves, including an image acquisition module, a preprocessing module, a multi-disease identification model module, and a decision-making and control module. Each module communicates in real time through a data transmission protocol, forming a closed-loop system from disease detection to precise control. The image acquisition module and the preprocessing module are connected via 5G or LoRa wireless transmission protocols to transmit the acquired cigar tobacco leaf images to the preprocessing module in real time. The preprocessing module is connected to the multi-disease identification model module via a data interface to input the processed image data into the multi-disease identification model module. The multi-disease identification model module is connected to the decision-making and control module via a communication protocol to transmit the disease identification results to the decision-making and control module. The decision-making and control module is connected to external control equipment via the MQTT protocol to realize the issuance and execution feedback of control commands.

[0005] The present invention further defines the technical solution as follows: Preferably, the image acquisition module uses a drone or a fixed field camera, wherein the drone is equipped with a multispectral camera that supports RGB + near-infrared imaging and has a resolution of not less than 2560×1440; the drone acquisition interval is 30 minutes / time; The fixed camera is a Hikvision DS-2CD3T86, which supports IP67 waterproofing, has 2 megapixels and night vision capabilities, takes a picture once per hour, and saves the captured image data in JPEG (RGB) and TIFF formats with a resolution of ≥2560×1440. It also records the shooting time, GPS coordinates, and temperature and humidity.

[0006] Preferably, the preprocessing module includes a data synthesis unit, a light correction unit, a dynamic background segmentation unit, and a data augmentation unit. The data synthesis unit generates lesion texture images using GAN. The light correction unit processes the images using the Retinex algorithm to eliminate uneven lighting and enhance the contrast between lesion areas and normal leaves. The dynamic background segmentation unit is based on a superpixel segmentation algorithm and integrates HSV color space and morphological filtering to extract leaf areas. The data augmentation unit processes the images through Mosaic synthesis and random rotation / scaling to simulate complex field scenarios such as leaf occlusion and angle changes. The preprocessed data is then transmitted to the multi-disease identification model module.

[0007] Preferably, the multi-disease identification model module includes a backbone network and multi-task branches. The backbone network uses a lightweight EfficientNet-B3 to extract multi-scale features. The multi-task branches include: a classification branch for simultaneously identifying disease types such as blue mold, black shank, red spot disease, and bacterial wilt; a detection branch for lesion location; and a severity assessment branch for calculating the lesion area ratio through regression analysis. The model embeds a CBAM attention mechanism to focus on lesion edges and texture features, improving identification accuracy. The processing results are sent to the decision-making and control module.

[0008] Preferably, the decision-making and control module includes a disease heat map generation unit, a control strategy generation unit, and an equipment linkage unit. The disease heat map generation unit generates a heat map by combining GPS coordinates and lesion distribution. The control strategy generation unit matches a pesticide database according to the disease type and severity and plans the agricultural machinery path based on the A* algorithm. The equipment linkage unit sends control commands to drones or smart sprayers via the MQTT protocol to achieve precise pesticide application and receives real-time pesticide application progress and abnormal alarm information from the equipment.

[0009] Preferably, the training strategy of the multi-disease identification model module is as follows: first, pre-train on the PlantVillage dataset for 100 rounds with a learning rate of 1e-3, and then fine-tune on a self-built cigar disease dataset containing 5000 labeled images and 500 synthetic images generated by GAN for 300 rounds with a learning rate of 1e-4. The input image size is uniformly 512×512.

[0010] Preferably, the prevention and control strategy generation unit of the decision-making and prevention and control module includes a rule engine, specifically: if the severity of blue mold is >15%, then control the drone to spray mancozeb at a dose of 20 ml / mu; if there is a mixed infection of red spot disease and bacterial wilt, then control the drone to spray difenoconazole and agricultural streptomycin at doses of 15 ml / mu and 10 ml / mu, respectively. The severity is determined according to GB / T 23222-2008 Classification and Investigation Methods for Tobacco Diseases and Pests.

[0011] The beneficial effects of this invention are: 1. This invention adopts a closed loop of image acquisition, preprocessing, recognition, decision-making, and control equipment linkage. Through the generation of disease heat maps (GPS + lesion distribution), A* algorithm planning of drone paths, and MQTT protocol real-time control of pesticide application equipment, it achieves seamless connection from detection to precise pesticide application, solving the problems of disconnect between decision-making and execution and excessive use of pesticides.

[0012] 2. This invention combines dynamic background segmentation (SLIC++ superpixel algorithm + HSV color space filtering) with illumination correction (Retinex algorithm) to specifically eliminate the effects of soil, stem interference, and uneven illumination, significantly improving the accuracy of lesion localization in scenarios with leaf shading, illumination changes, and soil interference. At the same time, it employs strategies such as GAN to generate synthetic data and Mosaic enhancement to simulate complex scenarios, significantly improving the model's adaptability to field environments.

[0013] 3. This invention uses the lightweight EfficientNet-B3 as the backbone network, integrating three branches: classification, detection, and severity assessment. It achieves simultaneous detection of blue mold, black shank, red spot disease, and bacterial wilt. By embedding the CBAM attention mechanism, it focuses on lesion features, solving the problems of insufficient adaptation of general models to cigar tobacco leaves, missed detection of small lesions (<5% leaf area), and detection conflicts among multiple diseases. It also avoids multiple calls to a single-task network, reducing processing time from >1 second / image in existing technologies to real-time levels (≤1 second). The missed detection rate for small lesions (<5% leaf area) is reduced from over 40% to an acceptable range, and the accuracy in actual field scenarios is improved to over 70%. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 The detection generated for this invention Figure 1 ; Figure 3 The detection generated for this invention Figure 2 ; Figure 4 The detection generated for this invention Figure 3 . Detailed Implementation Example

[0015] like Figure 1 As shown, this embodiment provides an image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves, including an image acquisition module, a preprocessing module, a multi-disease identification model module, and a decision-making and control module. Each module communicates in real time via a data transmission protocol, forming a closed-loop system from disease detection to precise control. The image acquisition module and the preprocessing module are connected via 5G or LoRa wireless transmission protocol to send the acquired cigar tobacco leaf images to the preprocessing module in real time; the preprocessing module and the multi-disease identification model module are connected via a data interface to input the processed image data into the multi-disease identification model module; the multi-disease identification model module and the decision-making and control module are connected via a communication protocol to transmit the disease identification results to the decision-making and control module; the decision-making and control module and external control equipment are connected via the MQTT protocol to realize the issuance and execution feedback of control commands.

[0016] The image acquisition module mentioned above uses a drone or a fixed field camera. The drone is equipped with a multispectral camera that supports RGB + near-infrared imaging and has a resolution of no less than 2560×1440. The drone acquisition interval is 30 minutes per acquisition. The fixed camera is a Hikvision DS-2CD3T86, which supports IP67 waterproofing, has 2 megapixels and night vision capabilities, takes a picture once per hour, and saves the captured image data in JPEG (RGB) and TIFF formats with a resolution of ≥2560×1440. It also records the shooting time, GPS coordinates, and temperature and humidity.

[0017] The aforementioned preprocessing module includes a data synthesis unit, a light correction unit, a dynamic background segmentation unit, and a data augmentation unit. The data synthesis unit generates lesion texture images using GAN. The light correction unit processes the images using the Retinex algorithm to eliminate uneven lighting and enhance the contrast between lesion areas and normal leaves. The dynamic background segmentation unit is based on a superpixel segmentation algorithm and integrates HSV color space and morphological filtering to extract leaf regions. The data augmentation unit synthesizes images using Mosaic and processes them by random rotation / scaling to simulate complex field scenarios such as leaf occlusion and angle changes. The preprocessed data is then transmitted to the multi-disease identification model module.

[0018] The aforementioned multi-disease identification model module includes a backbone network and multi-task branches. The backbone network uses a lightweight EfficientNet-B3 to extract multi-scale features. The multi-task branches include: a classification branch for simultaneously identifying disease types such as blue mold, black shank, red spot disease, and bacterial wilt; a detection branch for lesion location; and a severity assessment branch for calculating the lesion area ratio through regression analysis. The model incorporates a CBAM attention mechanism to focus on lesion edges and texture features, improving identification accuracy. The processing results are sent to the decision-making and control module.

[0019] The training strategy for the above-mentioned multi-disease identification model module is as follows: first, pre-train on the PlantVillage dataset for 100 rounds with a learning rate of 1e-3, and then fine-tune on a self-built cigar disease dataset containing 5000 labeled images and 500 synthetic images generated by GAN for 300 rounds with a learning rate of 1e-4. The input image size is uniformly 512×512.

[0020] The aforementioned decision-making and control module includes a disease heat map generation unit, a control strategy generation unit, and an equipment linkage unit. The disease heat map generation unit generates a heat map by combining GPS coordinates and lesion distribution. The control strategy generation unit matches the pesticide database according to the disease type and severity and plans the agricultural machinery path based on the A* algorithm. The equipment linkage unit sends control commands to drones or smart sprayers via the MQTT protocol to achieve precise pesticide application and receives real-time pesticide application progress and abnormal alarm information from the equipment.

[0021] The prevention and control strategy generation unit of the above-mentioned decision-making and prevention and control module includes a rule engine, specifically: if the severity of blue mold is >15%, then control the drone to spray mancozeb at a dose of 20 ml / mu; if there is a mixed infection of red spot disease and bacterial wilt, then control the drone to spray difenoconazole and agricultural streptomycin at doses of 15 ml / mu and 10 ml / mu, respectively. The severity is determined according to GB / T 23222-2008 Classification and Investigation Methods of Tobacco Diseases and Pests.

[0022] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A system for identifying and intelligently controlling multiple diseases in cigar tobacco leaves based on image detection, characterized in that: It includes an image acquisition module, a preprocessing module, a multi-disease identification model module, and a decision-making and control module. These modules communicate in real-time via a data transmission protocol, forming a closed-loop system from disease detection to precise control. The image acquisition module and the preprocessing module are connected via 5G or LoRa wireless transmission protocols to transmit the acquired cigar tobacco leaf images to the preprocessing module in real time. The preprocessing module is connected to the multi-disease identification model module via a data interface to input the processed image data into the multi-disease identification model module. The multi-disease identification model module is connected to the decision-making and control module via a communication protocol to transmit the disease identification results to the decision-making and control module. The decision-making and control module is connected to external control equipment via the MQTT protocol to realize the issuance and execution feedback of control commands.

2. The image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves according to claim 1, characterized in that: The image acquisition module uses a drone or a fixed field camera. The drone is equipped with a multispectral camera that supports RGB + near-infrared imaging and has a resolution of no less than 2560×1440. The drone acquisition interval is 30 minutes per acquisition. The fixed camera is a Hikvision DS-2CD3T86, which supports IP67 waterproofing, has 2 megapixels and night vision capabilities, takes a picture once per hour, and saves the captured image data in JPEG (RGB) and TIFF formats with a resolution of ≥2560×1440. It also records the shooting time, GPS coordinates, and temperature and humidity.

3. The image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves according to claim 2, characterized in that: The preprocessing module includes a data synthesis unit, a light correction unit, a dynamic background segmentation unit, and a data enhancement unit. The data synthesis unit generates lesion texture images using GAN. The light correction unit processes the images using the Retinex algorithm to eliminate uneven lighting and enhance the contrast between lesion areas and normal leaves. The dynamic background segmentation unit is based on a superpixel segmentation algorithm and integrates HSV color space and morphological filtering to extract leaf areas. The data augmentation unit simulates complex field scenarios such as leaf occlusion and angle changes by synthesizing images using Mosaic and processing them by random rotation / scaling. The pre-processed data is then transmitted to the multi-disease identification model module.

4. The image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves according to claim 3, characterized in that: The multi-disease identification model module includes a backbone network and multi-task branches. The backbone network uses a lightweight EfficientNet-B3 to extract multi-scale features. The multi-task branches include: a classification branch for simultaneously identifying disease types such as blue mold, black shank, red spot disease, and bacterial wilt; a detection branch for lesion location; and a severity assessment branch for calculating the lesion area ratio through regression analysis. The model incorporates a CBAM attention mechanism to focus on lesion edges and texture features, improving identification accuracy. The processing results are sent to the decision-making and control module.

5. The image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves according to claim 1, characterized in that: The decision-making and control module includes a disease heat map generation unit, a control strategy generation unit, and an equipment linkage unit. The disease heat map generation unit generates a heat map by combining GPS coordinates and lesion distribution. The control strategy generation unit matches the pesticide database according to the disease type and severity and plans the agricultural machinery path based on the A* algorithm. The equipment linkage unit sends control commands to the drone or smart sprayer through the MQTT protocol to achieve precise pesticide application and receives the pesticide application progress and abnormal alarm information of the equipment in real time.

6. The image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves according to claim 4, characterized in that: The training strategy for the multi-disease identification model module is as follows: first, it is pre-trained for 100 rounds on the PlantVillage dataset with a learning rate of 1e-3, and then fine-tuned for 300 rounds on a self-built cigar disease dataset containing 5000 labeled images and 500 synthetic images generated by GAN with a learning rate of 1e-4. The input image size is uniformly 512×512.

7. The image detection-based multi-disease identification and intelligent control system for cigar tobacco leaves according to claim 5, characterized in that: The prevention and control strategy generation unit of the decision-making and control module includes a rule engine, specifically: if the severity of blue mold is > 15%, then control the drone to spray mancozeb at a dose of 20 ml / mu; if there is a mixed infection of red spot disease and bacterial wilt, then control the drone to spray difenoconazole and agricultural streptomycin at doses of 15 ml / mu and 10 ml / mu, respectively. The severity is determined according to GB / T 23222-2008 Classification and Investigation Methods for Tobacco Diseases and Pests.

Images