An image recognition-based intelligent detection method and system for garden plant diseases and insect pests

Abstract

This application relates to the field of garden plant disease and pest detection technology, and provides an intelligent detection method and system for garden plant diseases and pests based on image recognition. The method includes: acquiring multispectral images of garden plants; determining whether liquid water exists on the leaf surface of the garden plants based on the reflectance of the multispectral images; when liquid water is present on the leaf surface of the garden plants, determining the type of liquid water in the multispectral images based on image features; processing the multispectral images based on the water type to eliminate optical distortion introduced by the liquid water, obtaining a corrected multispectral image; and judging plant diseases and pests based on the corrected multispectral image. This method can improve the accuracy of intelligent detection of garden plant diseases and pests.

Description

Technical Field

[0001] This application relates to the field of garden plant disease and pest detection technology, and in particular to an intelligent detection method and system for garden plant diseases and pests based on image recognition. Background Technology

[0002] In the daily health monitoring of garden plants, the use of image recognition technology, especially the combination of high-resolution multispectral imaging and deep learning methods, to intelligently detect plant diseases and pests has become an important development trend.

[0003] This method aims to overcome the shortcomings of traditional manual inspections, such as low efficiency and slow response. However, in practical applications, due to the complexity of the garden environment, especially when there is liquid water (such as water droplets or films) on the surface of plant leaves, existing technologies face severe challenges. These water bodies produce unique optical phenomena during image acquisition, such as the reflection and refraction of light, which visually resemble actual pathological changes in plants. This makes it difficult for intelligent recognition systems that rely on image features to accurately distinguish between them, resulting in frequent false alarms. This not only wastes a lot of manpower and resources on ineffective verification but also seriously affects users' trust in the system's reliability. Summary of the Invention

[0004] This application provides an intelligent detection method and system for garden plant diseases and pests based on image recognition, which can improve the accuracy of intelligent detection of garden plant diseases and pests.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] In a first aspect, this application discloses an intelligent detection method for garden plant diseases and pests based on image recognition, comprising: acquiring multispectral images of garden plants; determining whether liquid water exists on the leaf surface of garden plants based on the reflectance of the multispectral images; when liquid water exists on the leaf surface of garden plants, determining the type of liquid water in the multispectral images based on the image features of the multispectral images; processing the multispectral images based on the water type to eliminate the optical distortion introduced by the liquid water, obtaining a corrected multispectral image; and judging plant diseases and pests based on the corrected multispectral images.

[0007] This technical solution enables the effective identification of liquid water on the surface of garden plant leaves and allows for targeted correction based on the type of water, thereby eliminating the interference of water on image recognition, significantly improving the accuracy and reliability of pest and disease detection, and solving the problem of high false alarm rate caused by water in existing technologies.

[0008] Furthermore, in the above method, determining whether there is liquid water on the leaf surface of the garden plant based on the reflectance of the multispectral image includes: separating a first image in the near-infrared band and a second image in the green band from the multispectral image; and determining whether there is liquid water on the leaf surface of the garden plant based on the reflectance of the first image and the reflectance of the second image.

[0009] This technical solution utilizes the difference in sensitivity to water bodies between near-infrared and green light bands to more accurately detect the presence of liquid water on the leaf surface, providing accurate input for subsequent water body type identification and image correction.

[0010] Based on this, this application further proposes a method for determining whether liquid water exists on the surface of garden plant leaves based on the reflectance of a first image and the reflectance of a second image, comprising: using the ratio of the first value and the second value as a normalized differential water index; the first value being the difference between the reflectance of the first image and the reflectance of the second image, and the second value being the sum of the reflectance of the first image and the reflectance of the second image; and determining whether liquid water exists on the surface of garden plant leaves based on the normalized differential water index and the brightness value of the second image.

[0011] This application introduces a normalized differential water index, which, combined with brightness information in the green light band, can more robustly identify liquid water on the leaf surface, effectively distinguish water from the plant's own optical characteristics, and improve the accuracy of water detection.

[0012] More specifically, in some implementation schemes, determining whether liquid water exists on the surface of garden plant leaves based on the normalized difference water index and the brightness value of the second image includes: determining whether the normalized difference water index is less than a preset water index threshold; if the normalized difference water index is not less than the preset water index threshold, determining that liquid water does not exist on the surface of garden plant leaves; if the normalized difference water index is less than the preset water index threshold, determining whether the number of highlighted pixels in the second image is greater than a preset number threshold; if the brightness value of the highlighted pixels is greater than a preset brightness threshold; if the number of highlighted pixels in the second image is greater than the preset number threshold, determining that liquid water exists on the surface of garden plant leaves; otherwise, determining that liquid water does not exist on the surface of garden plant leaves.

[0013] By using this technical solution, this application can more accurately identify liquid water on the surface of a blade by setting a threshold and combining it with the judgment of the number of bright pixels, avoiding misjudgment that may be caused by a single indicator, and further improving the accuracy of water detection.

[0014] As an optional approach, in the above method, the image features include brightness values ​​and spectral curves. When there is liquid water on the surface of the leaves of garden plants, the water type of the liquid water in the multispectral image is determined based on the image features of the multispectral image, including: identifying areas in the multispectral image with brightness values ​​greater than a preset brightness threshold as liquid water based on the brightness values ​​of the multispectral image; and determining the water type of the liquid water based on the spectral curve of the liquid water.

[0015] This technical solution utilizes two image features—brightness value and spectral curve—to more comprehensively describe the optical properties of liquid water, thereby accurately distinguishing different types of water and laying the foundation for subsequent targeted correction.

[0016] To improve the scheme, the water body type includes water droplets or water film. The water body type of the liquid water body is determined based on the spectral curve of the liquid water body, including: determining the similarity between the spectral curve of the liquid water body and the preset drying spectral curve of the leaf surface in the multispectral image; if the similarity is less than the preset similarity threshold, the water body type of the liquid water body is determined to be water film; otherwise, the water body type of the liquid water body is determined to be water droplets.

[0017] This technical solution enables the effective differentiation of two common water body types—water droplets and water films—by comparing the similarity between the spectral curves of liquid water and those of dry water, providing a basis for subsequent accurate correction of different water body types.

[0018] Based on the above, this application further proposes that when the water body type is water droplets, the multispectral image is processed based on the water body type to eliminate the optical distortion introduced by the liquid water body and obtain a corrected multispectral image, including: calling a preset optical simulation model; obtaining the first brightness value of each pixel in the region corresponding to the water droplets; the first brightness value is determined according to the preset optical simulation model when there are no water droplets on the surface of the garden plant leaves; measuring the diameter of the liquid water body in the multispectral image; inputting the diameter of the liquid water body into the preset optical simulation model to obtain the second brightness value of each pixel in the region corresponding to the water droplets on the garden plant leaves; using the difference between the second brightness value and the first brightness value as the brightness value difference; using the difference between the current brightness value and the corresponding brightness value difference of each pixel in the region corresponding to the water droplets on the garden plant leaves in the multispectral image as the corrected brightness value, and obtaining the corrected multispectral image.

[0019] This technical solution addresses the optical distortion introduced by water droplets by using an optical simulation model to accurately calculate and compensate for brightness differences. This effectively eliminates the interference of water droplets on the image, restores the true optical information of the leaves, and significantly improves the accuracy of image correction.

[0020] As a technological improvement, when the water body type is a water film, the multispectral image is processed based on the water body type to eliminate the optical distortion introduced by the liquid water body and obtain a corrected multispectral image. This includes: obtaining a first preset correspondence; the first preset correspondence includes a one-to-one correspondence between multiple thickness ranges and multiple adjustment coefficients; determining the thickness of the water film based on the similarity between the spectral curve of the water film and the preset drying spectral curve of the leaf surface of garden plants when dry; using the adjustment coefficient corresponding to the thickness range of the water film in the first preset correspondence as the target adjustment coefficient; and using the product of the current spectral reflectance of the region corresponding to the water film in the multispectral image and the target adjustment coefficient as the corrected spectral reflectance to obtain the corrected multispectral image.

[0021] This technical solution addresses the optical distortion introduced by water film by establishing a correspondence between water film thickness and adjustment coefficient, and by performing spectral reflectance correction based on water film thickness. This effectively eliminates the interference of water film on the image, restores the true spectral information of the leaf, and improves the accuracy of image correction.

[0022] To further address the problem, the thickness of the water film is determined based on the similarity between the spectral curve of the water film and the preset drying spectral curve of the leaf surface of garden plants. This includes: obtaining a second preset correspondence; the second preset correspondence includes a one-to-one correspondence between multiple similarity ranges and multiple thicknesses; and using the thickness corresponding to the similarity range in the second preset correspondence as the thickness of the water film.

[0023] By establishing a correspondence between similarity and water film thickness, this application can more conveniently and accurately determine the thickness of the water film, simplifying the measurement process and improving the calibration efficiency.

[0024] Secondly, this application also discloses an intelligent detection system for garden plant diseases and pests based on image recognition, comprising: an acquisition device and a processing device; the acquisition device is used to acquire multispectral images of garden plants; the processing device is used to determine whether liquid water exists on the leaf surface of the garden plants based on the reflectance of the multispectral images; the processing device is used to determine the type of liquid water in the multispectral images based on image features when liquid water exists on the leaf surface of the garden plants; the processing device is used to process the multispectral images based on the water type to eliminate the optical distortion introduced by the liquid water and obtain a corrected multispectral image; the processing device is used to judge plant diseases and pests based on the corrected multispectral images.

[0025] Beneficial effects

[0026] This application discloses an intelligent detection method for garden plant diseases and pests based on image recognition. It acquires multispectral images of garden plants and determines the presence of liquid water on the leaf surface based on the image's reflectance. When liquid water is detected, the method further determines the specific type of water (e.g., water droplets or water film) based on image features and employs corresponding processing strategies to correct the multispectral image for different types of water, eliminating optical distortion introduced by the liquid water. Finally, plant disease and pest detection is made based on the corrected multispectral image. This method effectively solves the problem in existing technologies where liquid water (e.g., water droplets or water film) on the surface of garden plant leaves interferes with image recognition results, leading to misjudgments of diseases and pests. By identifying, classifying, and specifically correcting water bodies, this application significantly improves the accuracy and reliability of intelligent detection of garden plant diseases and pests, avoiding resource waste and reduced trust caused by false alarms, and providing more precise and efficient technical support for the health management of garden plants. Attached Figure Description

[0027] Figure 1 A flowchart illustrating an intelligent detection method for garden plant diseases and pests based on image recognition provided in this application;

[0028] Figure 2 A flowchart illustrating another intelligent detection method for garden plant diseases and pests based on image recognition provided in this application;

[0029] Figure 3 A flowchart illustrating another intelligent detection method for garden plant diseases and pests based on image recognition provided in this application;

[0030] Figure 4 This application provides a schematic diagram of the architecture of an intelligent detection system for garden plant diseases and pests based on image recognition. Detailed Implementation

[0031] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0032] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0033] Traditional intelligent detection methods for garden plant diseases and pests, when using image recognition technology for plant health monitoring, suffer from optical distortions due to the complexity of the garden environment, particularly the presence of liquid water (such as water droplets or films) on plant leaf surfaces. These distortions, such as light reflection and refraction, create visual similarities to actual pathological changes in plants. This makes it difficult for intelligent recognition systems that rely on image features to accurately distinguish between these phenomena and actual pathological changes. Consequently, the systems frequently issue false alarms, wasting significant human and material resources on ineffective verification and severely impacting user trust in the system's reliability.

[0034] In this regard, such as Figure 1 As shown, this application proposes an intelligent detection method for garden plant diseases and pests based on image recognition, including:

[0035] S101. Obtain multispectral images of garden plants.

[0036] S102. Determine whether there is liquid water on the surface of garden plant leaves based on the reflectance of multispectral images.

[0037] S103. When there is liquid water on the surface of the leaves of garden plants, determine the type of liquid water in the multispectral image based on the image features of the multispectral image.

[0038] S104. Process the multispectral image based on the water body type to eliminate the optical distortion introduced by liquid water and obtain the corrected multispectral image.

[0039] S105. Based on the corrected multispectral images, determine plant diseases and pests.

[0040] This application utilizes the aforementioned method to effectively identify and eliminate optical distortions caused by liquid water in multispectral images, thereby improving the accuracy and reliability of plant disease and pest detection, reducing false alarms, and enhancing the practical value of the system.

[0041] To better understand the technical solutions proposed in this application, the following will explain some key terms and implementation environments involved. "Multispectral image" as mentioned in this application refers to image data containing multiple narrow-band spectral information acquired by devices such as multispectral cameras. These bands typically cover the visible light and near-infrared ranges, providing richer plant physiological information than traditional RGB images. "Reflectivity" refers to the ratio of light energy reflected from an object's surface to incident light energy, a crucial characteristic of pixels in a multispectral image, reflecting the object's absorption and reflection characteristics of different wavelengths of light. "Liquid water" specifically refers to liquid water present on the surface of garden plant leaves, including but not limited to water droplets and water films, which alter the optical properties of the leaf surface. "Image features" refers to quantitative indicators used to describe the content or regional characteristics of an image, such as brightness values, spectral curves, and texture features. The implementation environment of this application typically includes multispectral image acquisition equipment (such as multispectral cameras mounted on drones, handheld multispectral analyzers, etc.), image processing units (such as computers, embedded processors, etc.), and corresponding software algorithms.

[0042] The core of the intelligent detection method for garden plant diseases and pests based on image recognition proposed in this application lies in the identification and correction of optical distortion introduced by liquid water.

[0043] First, it is necessary to acquire multispectral images of the garden plants. Multispectral image acquisition can be achieved in several ways. For example, a drone equipped with a multispectral camera can be used for large-scale, high-efficiency image acquisition. The drone flies along a preset route and altitude, periodically taking images of the garden plants. Another method is to use a handheld multispectral instrument, where an operator takes close-up photos of the target plants from the ground. This method is suitable for detailed inspection of specific plants or localized areas. Alternatively, a fixed multispectral sensor array can be used to continuously monitor garden plants in a specific area to obtain time-series multispectral image data.

[0044] Secondly, the presence of liquid water on the leaf surface of garden plants can be determined based on the reflectance of multispectral images. One approach is to analyze the reflectance characteristics of different bands in the multispectral image. For example, specific bands sensitive to water (such as near-infrared and green light bands) can be selected, and the differences or ratios of reflectance in these bands can be calculated. When liquid water is present on the leaf surface, its reflectance in these bands will change significantly; for example, near-infrared reflectance decreases, while green light reflectance relatively increases. By setting appropriate thresholds, the presence of liquid water can be determined. Another approach utilizes a machine learning model. By collecting a large number of multispectral image samples containing and not containing liquid water, and labeling these samples, a classification model can be trained. This model can learn the reflectance patterns of liquid water in multispectral images, thereby determining the presence of liquid water in new images.

[0045] Furthermore, when liquid water is present on the surface of garden plant leaves, the type of liquid water in the multispectral image can be determined based on its image features. For example, image features such as brightness values ​​and spectral curves of the liquid water area can be extracted. Water droplets and water films exhibit different brightness distributions and spectral responses in multispectral images due to their different morphologies and optical properties. Water droplets typically appear as bright, well-defined circular or elliptical areas with uneven brightness distribution and potential highlights. Water films, on the other hand, may appear as relatively uniformly bright areas with a larger coverage area, and their spectral curves may differ from those of dry leaves. By analyzing these characteristics, water droplets and water films can be distinguished.

[0046] Next, the multispectral image is processed based on the water body type to eliminate the optical distortion introduced by liquid water, resulting in a corrected multispectral image. Different correction strategies are required for different water body types. For example, water droplets, due to their lensing effect, can cause light to focus or scatter in the area below the leaf, leading to abnormal local brightness. In this case, an optical model can be established to simulate the effect of water droplets on light, and a correction factor can be calculated based on parameters such as the size and shape of the water droplets to adjust the pixel brightness of the affected area. For water films, because they form a thin layer covering the leaf surface, they change the overall reflectivity of the leaf surface. In this case, a reflectivity correction model can be established based on the thickness and optical characteristics of the water film to adjust the spectral reflectivity of the water film-covered area to restore the true reflectivity of the leaf.

[0047] Finally, plant diseases and pests are identified based on the corrected multispectral images. After obtaining the corrected multispectral images, various image recognition techniques can be used for disease and pest identification. For example, deep learning models, such as convolutional neural networks (CNNs), can be used to analyze the corrected images. These models, trained on a large number of disease and pest image samples, can learn and identify the unique image features of different diseases and pests, thereby accurately determining whether plants are diseased and the type of disease. In addition, traditional image processing methods, such as texture analysis, color analysis, and shape analysis, can be combined to extract features of diseased areas and compare them with a pre-set disease feature database for disease and pest classification and identification.

[0048] The intelligent detection method for garden plant diseases and pests based on image recognition proposed in this application effectively solves the problem of decreased detection accuracy of traditional methods in complex garden environments due to liquid water on the leaf surface, by introducing liquid water body recognition, water body type judgment and targeted optical distortion correction mechanism.

[0049] Specifically, the core innovations of this application are as follows: First, by analyzing the reflectance of multispectral images, it is possible to accurately determine whether there is liquid water on the blade surface, avoiding unnecessary corrections in the absence of water interference and improving processing efficiency. Second, it further distinguishes the types of liquid water (such as water droplets or water films) and adopts different optical distortion correction strategies for different types of water bodies, making the correction process more targeted and accurate. For example, for water droplets, the local brightness anomalies caused by their lens effect can be finely corrected using an optical simulation model; while for water films, their impact on overall reflectance is compensated by adjusting coefficients. This meticulous classification process is significantly superior to the existing technology's approach of treating all liquid water bodies individually or coarsely.

[0050] Compared to the closest existing technology, the advantage of this application lies in its refined processing of environmental factors. Existing technologies often directly determine pests and diseases from multispectral images, ignoring the potential impact of liquid water on image quality, leading to a high false alarm rate in humid environments such as after rain or with dew. This application effectively eliminates these interfering factors by adding liquid water detection and optical distortion correction steps before pest and disease assessment, enabling subsequent assessments to be based on more realistic and accurate plant image information. Therefore, this application can significantly improve the accuracy and reliability of intelligent detection of garden plant pests and diseases, reduce the additional manpower and material resources consumed due to false alarms, and enhance user trust in the intelligent detection system, thus possessing significant practical application value.

[0051] Specifically, in the above-mentioned intelligent detection method for garden plant diseases and pests based on image recognition, the following method can be used to determine whether there is liquid water on the surface of garden plant leaves based on the reflectance of multispectral images.

[0052] like Figure 2 As shown, according to the above method, determining whether liquid water exists on the surface of garden plants' leaves based on the reflectance of multispectral images includes:

[0053] S201. Separate the first image of the near-infrared band and the second image of the green band from the multispectral image.

[0054] S202. Determine whether there is liquid water on the surface of the leaves of the garden plants based on the reflectance of the first image and the reflectance of the second image.

[0055] Specifically, multispectral images typically contain spectral information across multiple different bands. To more accurately detect liquid water on leaf surfaces, specific bands sensitive to water can be extracted from these multispectral data. The first image, in the near-infrared band, refers to image data acquired within the near-infrared spectral range. This band has strong penetrating and reflective properties for understanding plant internal structure and water content. The second image, in the green light band, refers to image data acquired within the green light spectral range. This band is typically strongly absorbed by plant chlorophyll but also responds to some water reflectance. By separating these two specific bands, more targeted spectral information can be provided for subsequent water detection.

[0056] The reflectance of the first image and the reflectance of the second image refer to the intensity of light reflected from the blade surface within their respective spectral bands. Liquid water, such as dew or raindrops, exhibits unique reflective properties in different spectral bands. For example, water has a high absorptivity and low reflectivity in the near-infrared band, while its reflectivity is relatively high in the green light band. By comparing and analyzing the reflectance in these two bands, the presence of liquid water on the blade surface can be effectively identified.

[0057] This application's method separates a first image in the near-infrared band and a second image in the green band from a multispectral image, and determines the presence of liquid water on the leaf surface based on the reflectance of these two specific bands. The working principle is that water has significantly different optical properties in different spectral bands. Specifically, liquid water absorbs near-infrared light much more than green light, resulting in a significantly reduced reflectance in the near-infrared band and a relatively high reflectance in the green band. By acquiring and analyzing the reflectance difference between these two bands, a water-sensitive index can be constructed, thereby achieving effective detection of liquid water on the leaf surface. This analytical method based on reflectance in specific spectral bands can utilize the unique spectral fingerprint of water, improving the accuracy and reliability of water detection.

[0058] The above technical solution utilizes the unique spectral response differences of water in the near-infrared and green light bands to more accurately identify the presence of liquid water on the surface of garden plant leaves. This method avoids potential misjudgments that may arise from relying solely on a single band or panspectral information, improves the sensitivity and specificity of liquid water detection, and provides more accurate input data for subsequent optical distortion correction and pest and disease assessment.

[0059] This application further proposes a more accurate method for determining the existence of liquid water by introducing a normalized differential water index and the brightness value of a second image for comprehensive judgment.

[0060] like Figure 3 As shown, this application further proposes the following steps for determining whether liquid water exists on the surface of garden plant leaves based on the reflectance of the first image and the reflectance of the second image:

[0061] S301. The ratio of the first value to the second value is used as the normalized differential water index.

[0062] The first value is the difference between the reflectance of the first image and the reflectance of the second image, and the second value is the sum of the reflectance of the first image and the reflectance of the second image.

[0063] S302. Determine whether there is liquid water on the leaf surface of garden plants based on the normalized differential water index and the brightness value of the second image.

[0064] Specifically, the first value refers to the difference between the reflectance of the first image and the reflectance of the second image, i.e., NIR-Green. The second value refers to the sum of the reflectance of the first image and the second image, i.e., NIR+Green. The Normalized Difference Water Index (NDWI) is the ratio of these two values, i.e., (NIR-Green) / (NIR+Green). This index is widely used for water body identification. Its principle is that water has low reflectance in the near-infrared band and relatively high reflectance in the green light band. Therefore, the NDWI value of water areas is usually positive and high, or in some cases, negative but within a specific range. The brightness value of the second image can be understood as the pixel intensity of the green light band image. In practical applications, liquid water, especially water droplets or films, usually exhibits high brightness in the green light band due to the reflection or transmission characteristics of green light. Therefore, combining the brightness value in the green light band can further assist in determining the presence of liquid water.

[0065] This application's solution, by introducing a normalized differential water index and the brightness value of a second image, can more effectively identify liquid water on the surface of garden plant leaves. The normalized differential water index utilizes the difference in reflectance characteristics of water in the near-infrared and green light bands, giving water areas a unique index value in the image, thereby enhancing the contrast between water and non-water areas. Simultaneously, combining this with the brightness value of the second image can further verify and assist in the judgment. For example, when the normalized differential water index indicates the possible presence of water, if the brightness value in the green light band is also high, the reliability of the judgment is further enhanced, as water typically appears as a bright area in the green light band. This dual judgment mechanism effectively avoids misjudgments that may occur with a single reflectance judgment, improving the accuracy and robustness of water detection.

[0066] Through the above technical solution, this application can more accurately determine whether there is liquid water on the surface of garden plant leaves. The introduction of the normalized difference water index means that water identification no longer relies solely on the reflectance of a single wavelength band, but utilizes the unique spectral response differences of water bodies across different wavelength bands, thereby improving the sensitivity and specificity of water body detection. Furthermore, combining the brightness value of a second image for auxiliary judgment further enhances the ability to distinguish high-reflectance non-water areas, effectively reducing the false positive rate. Therefore, this application can more accurately identify liquid water on leaf surfaces under complex environmental conditions, providing more reliable basic data for subsequent optical distortion elimination and pest and disease assessment.

[0067] In some preferred embodiments, specifically, after acquiring a multispectral image of the garden plants, a first image in the near-infrared band and a second image in the green band are first separated from the multispectral image. Then, for each pixel in the image, its normalized difference water index is calculated. For example, if a pixel has a reflectance of 0.1 in the first image and a reflectance of 0.3 in the second image, its first value is 0.1-0.3=-0.2, its second value is 0.1+0.3=0.4, and the normalized difference water index is -0.2 / 0.4=-0.5. Simultaneously, the brightness value of the pixel in the second image is obtained. By comprehensively analyzing these calculated normalized difference water indices and the brightness values ​​of the second image, it can be determined whether liquid water exists on the leaf surface corresponding to the pixel. For example, a normalized difference water index threshold and a brightness value threshold can be set. When the normalized difference water index of a pixel is less than the preset water index threshold (for example, a negative value, indicating that the water absorbs near-infrared light and reflects green light) and its green light brightness value is higher than the preset brightness threshold, it is determined that there is liquid water in the area.

[0068] In some of the embodiments described above in this application, a method is proposed for determining whether liquid water exists on the surface of the leaves of garden plants based on the reflectance of multispectral images.

[0069] Specifically, the process of determining whether liquid water exists on the leaf surface of garden plants based on the normalized difference water index and the brightness value of the second image can be further refined into the following steps:

[0070] Determine if the normalized difference water index is less than a preset water index threshold; if the normalized difference water index is not less than the preset water index threshold, determine that there is no liquid water on the leaf surface of the garden plant; if the normalized difference water index is less than the preset water index threshold, determine if the number of highlighted pixels in the second image is greater than a preset number threshold; if the brightness value of the highlighted pixels is greater than a preset brightness threshold; if the number of highlighted pixels in the second image is greater than the preset number threshold, determine that there is liquid water on the leaf surface of the garden plant; otherwise, determine that there is no liquid water on the leaf surface of the garden plant.

[0071] The Normalized Difference Water Index (NDWI) is a commonly used remote sensing index. Its calculation is typically based on the reflectance of a first image in the near-infrared band and the reflectance of a second image in the green band. This index aims to highlight water bodies in the image; areas with water bodies generally show lower NDWI values. A preset NDWI threshold is a critical value set based on experience or experimental data to preliminarily determine the presence of water bodies. When the NDWI is lower than this preset threshold, it indicates that liquid water may exist in the area.

[0072] Furthermore, the second image refers to the green light band image separated from the multispectral image. High-brightness pixels are those in the second image whose brightness value exceeds a preset brightness threshold. The preset brightness threshold is used to distinguish between normal leaf areas and high-brightness areas that may be caused by water reflection. The preset quantity threshold is used to determine whether the number of high-brightness pixels is sufficient to constitute a water area, rather than accidental noise or reflective points.

[0073] This application's solution, by combining the Normalized Difference Water Index (NDDI) and the brightness value of a second image, can more accurately identify liquid water bodies on the surface of garden plant leaves. Specifically, liquid water bodies exhibit strong absorption characteristics in the near-infrared band and certain reflective characteristics in the green light band, which typically results in a low NDDI for water bodies. Therefore, by first determining whether the NDDI is less than a preset water index threshold, potential water body areas can be preliminarily screened.

[0074] However, relying solely on the Normalized Difference Water Index (NDDI) may not be sufficient to completely distinguish water bodies from surfaces with low NDDI values, such as certain shaded areas or soil. Therefore, this solution further incorporates an assessment of the number of bright pixels in the second image. When water droplets or films are present on leaf surfaces, especially under sunlight, water bodies may form bright reflective points or areas. These bright pixels appear as high brightness values ​​in the second image in the green light band. By determining whether the number of these bright pixels exceeds a preset threshold, low NDDI values ​​caused by non-water body factors can be effectively excluded, thereby improving the accuracy of liquid water detection.

[0075] Through the above technical solution, this application can perform preliminary screening using the normalized difference water index's sensitivity to water bodies, and then perform secondary verification by combining this with the high-brightness reflection phenomenon unique to water bodies in green light band images. This two-stage judgment mechanism effectively avoids misjudgments that may be caused by a single indicator, such as misidentifying shadows or soil as water bodies, or omitting water bodies with atypical normalized difference water indexes due to complex lighting conditions. Therefore, it improves the accuracy and robustness of detecting liquid water on the surface of garden plant leaves, providing more reliable input data for subsequent optical distortion elimination and pest and disease assessment.

[0076] Specifically, in some of the above embodiments, in order to more accurately determine the water body type of liquid water in multispectral images, this application provides a detailed description of the image features used and gives a specific method for determining the water body type based on these features.

[0077] Image features include brightness values ​​and spectral curves. When liquid water is present on the surface of garden plant leaves, the type of liquid water in the multispectral image is determined based on the image features, including:

[0078] Based on the brightness values ​​of the multispectral images, regions with brightness values ​​greater than a preset brightness threshold are identified as liquid water bodies; the water body type of the liquid water body is determined based on its spectral curve.

[0079] Specifically, image features refer to the brightness values ​​and spectral curves of multispectral images. Brightness values ​​reflect the lightness or darkness of pixels in an image, while spectral curves describe the reflection or absorption characteristics of an object at different wavelengths. In particular, when identifying liquid water bodies, preliminary screening can be performed based on the brightness values ​​of multispectral images. Typically, liquid water bodies, especially water droplets or films, exhibit higher brightness values ​​in multispectral images due to the difference in their light reflection characteristics compared to plant leaves. Therefore, areas in multispectral images with brightness values ​​exceeding a preset brightness threshold can be identified as potential liquid water bodies.

[0080] Furthermore, after identifying potential liquid water regions, the specific type of liquid water can be determined by analyzing the spectral curves of these regions. Different types of water bodies (such as water droplets and water films) exhibit different spectral characteristics in terms of reflectance or absorptivity at different wavelengths. By comparing the spectral curves of liquid water regions with standard spectral curves of known water types, their water body type can be accurately determined.

[0081] This application's solution utilizes both brightness values ​​and spectral curves from multispectral images to more precisely identify and differentiate liquid water bodies on the surface of garden plant leaves. Brightness values, as an intuitive image feature, can quickly locate potential water bodies, avoiding complex spectral analysis of the entire image and thus improving processing efficiency. Spectral curves, on the other hand, provide deeper physical information, revealing the specific morphology of the water body (such as water droplets or films), which is difficult to achieve using brightness values ​​alone. This step-by-step and complementary feature utilization method lays the foundation for subsequent precise optical distortion removal for different water body types, ensuring the accuracy of pest and disease assessment.

[0082] Through the above technical solution, this application can effectively distinguish different types of liquid water bodies on the surface of garden plant leaves. This distinction is crucial for subsequently eliminating the optical distortion introduced by liquid water bodies, because different types of water bodies (such as water droplets and water films) exhibit significant differences in the refraction, reflection, and scattering effects of light, requiring different correction strategies. By accurately identifying the water body type, the most suitable correction method can be selected to more precisely eliminate optical distortion, improve the quality of the corrected image, and ultimately enhance the accuracy and reliability of image recognition-based intelligent detection of garden plant diseases and pests.

[0083] In the aforementioned intelligent detection method for garden plant diseases and pests based on image recognition, the water body type includes water droplets or water films. The water body type is determined based on the spectral curve of the liquid water body, including:

[0084] Determine the similarity between the spectral curve of the liquid water body and the preset drying spectral curve of the leaf surface in the multispectral image; if the similarity is less than the preset similarity threshold, determine the water body type of the liquid water body as a water film; otherwise, determine the water body type of the liquid water body as a water droplet.

[0085] Specifically, the aforementioned water body types can be understood as the specific forms in which liquid water exists on the surface of garden plant leaves, mainly including two types: water droplets and water films. Water droplets usually refer to separate droplets with a certain curvature formed on the leaf surface, while water films refer to a thin layer of water that is evenly spread on the leaf surface.

[0086] The similarity determination between the spectral curve of the liquid water body and the preset dry spectral curve of the leaf surface in the multispectral image refers to the degree of matching between the spectral curve of the liquid water body region and the preset dry spectral curve representing the spectral characteristics of the plant leaf in a completely dry state. This similarity can be calculated using various mathematical methods, such as Euclidean distance, cosine similarity, or correlation coefficient, with the aim of quantifying the influence of liquid water on the spectral reflectance characteristics of the leaf. The preset dry spectral curve is a set of reflectance values ​​of a specific garden plant's leaves in different wavelength bands under a waterless state, which can be obtained through experimental measurements or database queries.

[0087] Furthermore, the preset similarity threshold is a key parameter used to distinguish between water droplets and water films. Its value can be obtained based on experimental data, empirical values, or through training and optimization using machine learning methods. When the calculated similarity is less than the preset similarity threshold, it indicates that the liquid water body has a greater impact on the spectral characteristics of the leaf, and it is more likely to be identified as a water film; conversely, when the similarity is not less than the threshold, it is considered that the liquid water body has a relatively smaller impact on the spectral characteristics of the leaf, and it is more likely to be identified as a water droplet.

[0088] This application's solution distinguishes between water droplets and water films by comparing the similarity between the spectral curves of liquid water regions and preset dry spectral curves. The working principle is as follows: Water droplets, due to their lensing effect, introduce optical distortion, but the spectral characteristics of the underlying leaf can still be detected to some extent. Therefore, their spectral curves have a relatively high similarity to those of dry leaves. Water films, on the other hand, more significantly alter the optical properties of the leaf surface, resulting in a greater difference in spectral reflectance between them and dry leaves, thus leading to a relatively low similarity. It is precisely because water droplets and water films affect the spectral reflectance characteristics of leaves through different mechanisms that similarity comparison can effectively distinguish them.

[0089] Through the above technical solution, this application can accurately distinguish whether the liquid water on the surface of garden plant leaves is a droplet or a film, based on the degree of influence of liquid water on the spectral reflectance characteristics of leaves. This refined water type identification allows for more targeted correction methods and parameters to be used for optical distortions introduced by different water types. For example, water droplets can be corrected using an optical simulation model, while water films can be corrected using adjustment coefficients. This significantly improves the correction accuracy of multispectral images, providing more accurate image data for subsequent plant disease and pest assessment, thereby enhancing the overall accuracy and reliability of disease and pest detection.

[0090] In some preferred embodiments, a specific example is given below. Suppose that pest and disease detection is needed on the leaves of a specific rose variety. First, in a waterless state, an image of the rose leaf is acquired using a multispectral camera, and its spectral curve is extracted and used as a preset dry spectral curve. Subsequently, when a multispectral image of the rose leaf containing liquid water is acquired, the liquid water region is first identified based on its brightness value, and the spectral curve of that region is extracted. Next, the cosine similarity between the spectral curve of the liquid water and the preset dry spectral curve is calculated. For example, if experiments determine that a similarity less than 0.8 indicates a water film, and otherwise a water droplet, then 0.8 is set as the preset similarity threshold. If the calculated similarity is 0.75, the liquid water is determined to be a water film; if the calculated similarity is 0.85, the liquid water is determined to be a water droplet. In this way, the specific form of liquid water on the leaf surface can be accurately identified, providing crucial information for subsequent image correction.

[0091] This application further proposes a specific method for processing multispectral images based on the water body type when the water body type is water droplets, in order to eliminate the optical distortion introduced by liquid water and obtain a corrected multispectral image.

[0092] Specifically, when the water body type is water droplets, the multispectral image is processed based on the water body type to eliminate the optical distortion introduced by the liquid water body, resulting in a corrected multispectral image, including:

[0093] Call the preset optical simulation model; obtain the first brightness value of each pixel in the region corresponding to the water droplet; the first brightness value is the brightness value of each pixel in the region corresponding to the water droplet on the leaf of the garden plant when there is no water droplet on the leaf surface of the garden plant, determined according to the preset optical simulation model; measure the diameter of the liquid water body in the multispectral image; input the diameter of the liquid water body into the preset optical simulation model to obtain the second brightness value of each pixel in the region corresponding to the water droplet on the leaf of the garden plant; take the difference between the second brightness value and the first brightness value as the brightness value difference; take the difference between the current brightness value of each pixel in the region corresponding to the water droplet on the leaf of the garden plant in the multispectral image and the corresponding brightness value difference as the corrected brightness value, and obtain the corrected multispectral image.

[0094] Calling the preset optical simulation model refers to activating a pre-established model that can simulate the interaction between light and the surface of a plant leaf. This model can predict the spectral response of a specific region on the leaf surface in the presence of water droplets, based on physical optics principles such as refraction, reflection, and scattering. Its purpose is to provide a theoretical basis and simulation environment for subsequent brightness value calculations.

[0095] Obtaining the first brightness value of each pixel in the region corresponding to the water droplet refers to calculating the brightness value that each pixel should have in the region originally covered by the water droplet using the aforementioned preset optical simulation model, under the ideal condition that there are no water droplets on the surface of the garden plant's leaves. This first brightness value represents the true spectral information of the leaf under the condition of no water droplet interference and serves as the benchmark for optical distortion correction.

[0096] Measuring the diameter of liquid water in a multispectral image refers to accurately identifying and measuring the geometric dimensions, particularly the diameter, of a liquid water droplet from an acquired multispectral image using image processing techniques such as edge detection and morphological manipulation. The diameter of a water droplet is a key parameter affecting its optical properties, such as its ability to focus or scatter light.

[0097] By inputting the diameter of the liquid water body into a preset optical simulation model, the second brightness value of each pixel in the region corresponding to the water droplet on the leaf of a garden plant is obtained. This means that the actual measured diameter of the water droplet is input as a parameter into the preset optical simulation model to simulate the brightness value of each pixel in the area covered by the water droplet when a water droplet of that specific diameter is present. This second brightness value reflects the brightness performance of the water droplet after optical distortion is introduced.

[0098] The difference between the second brightness value and the first brightness value is used as the brightness value difference, which refers to the difference between the calculated and simulated brightness value with water droplets (the second brightness value) and the brightness value without water droplets (the first brightness value). This brightness value difference quantifies the specific impact of water droplets on the brightness of the leaf surface, that is, the degree of optical distortion introduced by the water droplets.

[0099] The corrected multispectral image is obtained by subtracting the difference between the current brightness value and the corresponding brightness value of each pixel in the region corresponding to water droplets on the leaves of garden plants in the actual multispectral image. This method extracts the distortion introduced by water droplets from the observed brightness value, resulting in a corrected brightness value that more closely approximates the true spectral reflectance of the leaves, ultimately forming a corrected multispectral image free from optical distortion.

[0100] This application's solution, by introducing a pre-defined optical simulation model, can accurately simulate the effect of water droplets on light on the leaf surface. By acquiring a first brightness value without water droplets and a second brightness value with water droplets (of a specific diameter), the brightness change introduced by the water droplets can be quantified, i.e., the brightness difference. Subtracting this brightness difference from the current brightness value of the water droplet region in the multispectral image effectively inverts and eliminates the optical distortion caused by the water droplets, thereby obtaining a corrected image that more closely approximates the spectral information of the real leaf surface.

[0101] Through the above technical solution, this application provides a high-precision optical distortion correction method for water droplets, a specific type of liquid water. Compared with solutions that do not distinguish between water types or use general correction methods, this application can more accurately eliminate the interference of water droplets on multispectral images, significantly improve the quality of the corrected image and the accuracy of subsequent plant disease and pest assessment, especially in scenarios where the optical effects of water droplets are significant.

[0102] In some preferred embodiments, assuming that after acquiring a multispectral image of the garden plants, analysis determines the presence of liquid water on the leaf surface, and the water type is identified as water droplets, the system invokes a pre-trained optical simulation model. This model can predict the impact of water droplets on the spectral reflectance of the leaf surface based on the diameter of the water droplets and the incident light conditions. First, the model is used to calculate the first brightness value that each pixel in the water droplet-covered area should have in the absence of water droplets. Next, the diameter of the water droplets in the actual multispectral image is measured using image processing techniques, and this diameter is input into the optical simulation model to simulate the second brightness value that each pixel in the water droplet-covered area should have when water droplets of that diameter are present. Then, the difference between the second brightness value and the first brightness value is calculated to obtain the brightness deviation of each pixel due to the presence of water droplets. Finally, the current brightness value of each pixel in the water droplet-covered area of ​​the multispectral image is subtracted from the corresponding brightness deviation to obtain the corrected brightness value, forming a corrected multispectral image that eliminates the optical distortion caused by water droplets, for subsequent pest and disease assessment.

[0103] This application further proposes a method for processing the multispectral image based on the water body type when the water body is a water film, in order to eliminate the optical distortion introduced by the liquid water body and obtain a corrected multispectral image, the steps of which include:

[0104] Obtain a first preset correspondence; the first preset correspondence includes a one-to-one correspondence between multiple thickness ranges and multiple adjustment coefficients; determine the thickness of the water film based on the similarity between the spectral curve of the water film and the preset drying spectral curve of the leaf surface of the garden plant when it is dry; take the adjustment coefficient corresponding to the thickness range of the water film in the first preset correspondence as the target adjustment coefficient; take the product of the current spectral reflectance of the area corresponding to the water film in the multispectral image and the target adjustment coefficient as the corrected spectral reflectance, and obtain the corrected multispectral image.

[0105] Specifically, the first preset correspondence refers to a pre-established mapping table or function that describes the relationship between water film thickness and optical adjustment coefficients. This correspondence can be obtained through experimental measurement, physical modeling, or machine learning, with the aim of quantifying the impact of water film thicknesses on spectral reflectance and providing corresponding correction parameters. Multiple thickness ranges can divide the possible thicknesses of the water film into several intervals, each interval corresponding to one or a set of adjustment coefficients to achieve refined correction.

[0106] Furthermore, the similarity between the spectral curve of the water film and the preset drying spectral curve of the leaf surface of garden plants can be calculated using various spectral similarity measurement methods, such as spectral angle mapping (SAM), spectral correlation coefficient, and Euclidean distance. The preset drying spectral curve is a standard spectral reflectance curve of the garden plant leaf surface under conditions without water film coverage, which is collected or modeled in advance and serves as a reference. By comparing the similarity between the actual spectral curve of the water film area and this drying spectral curve, the degree of absorption and scattering of the spectrum by the water film can be indirectly reflected, thereby inferring the thickness of the water film.

[0107] Therefore, after determining the thickness of the water film, the thickness range within which that thickness falls can be found according to the first preset correspondence, and the corresponding target adjustment coefficient can be obtained. This target adjustment coefficient is designed for the optical properties of a water film of a specific thickness and is used to compensate for the influence of the water film on the spectral reflectance.

[0108] Finally, the corrected spectral reflectance is obtained by multiplying the current spectral reflectance of the region corresponding to the water film in the multispectral image with the acquired target adjustment coefficient. This corrected spectral reflectance is closer to the true spectral response of the blade surface without water film coverage, thus eliminating the optical distortion introduced by the water film.

[0109] This application's solution establishes a correspondence between water film thickness and an optical adjustment coefficient, and indirectly determines the water film thickness by utilizing the similarity between the water film's spectral curve and the spectral curve of a dried leaf, thereby effectively eliminating the optical distortion introduced by the water film. Specifically, the presence of a water film alters the spectral reflectance characteristics of the leaf surface, and this alteration is closely related to the water film's thickness. When the water film is thin, its absorption and scattering of the spectrum are weaker; when the water film is thicker, its effect is more significant. By comparing the spectral curves of the water film region with those of the dried leaf, this difference can be quantified and matched with a pre-established similarity-thickness relationship, thereby accurately estimating the water film thickness. Once the water film thickness is determined, a target adjustment coefficient for that thickness can be found from a first pre-defined correspondence. This adjustment coefficient is derived based on extensive experimental data or physical models and can accurately compensate for the attenuation or enhancement effects of a water film of a specific thickness on spectral reflectance. By multiplying the current spectral reflectance by the target adjustment coefficient, the spectral data affected by the water film can be restored to a spectral response close to that of the actual leaf surface, thereby eliminating the optical distortion caused by the presence of the water film and providing more accurate image data for subsequent disease and pest assessment.

[0110] Through the above technical solution, this application provides a refined optical distortion correction method for situations where a water film exists on the surface of garden plant leaves. This method effectively eliminates the optical distortion introduced by the water film by considering the influence of water film thickness on spectral reflectance and using a preset thickness-adjustment coefficient correspondence for correction. Compared to solutions without water film correction or using general correction methods, this application can obtain corrected multispectral images that more closely resemble the true spectral characteristics of leaves, significantly improving the accuracy and reliability of subsequent plant disease and pest assessment, avoiding misjudgments or missed detections caused by water film interference, thereby enhancing the overall performance of intelligent detection of garden plant diseases and pests.

[0111] Specifically, in some of the above embodiments, the thickness of the water film is determined based on the similarity between the spectral curve of the water film and the preset drying spectral curve of the leaf surface of the garden plant when it is dry. This can be further achieved in the following ways.

[0112] The thickness of the water film is determined based on the similarity between its spectral curve and a preset drying spectral curve of the leaf surface of the garden plant. This includes:

[0113] Obtain a second preset correspondence; the second preset correspondence includes a one-to-one correspondence between multiple similarity ranges and multiple thicknesses; the thickness corresponding to the similarity range in the second preset correspondence is taken as the thickness of the water film.

[0114] The second pre-defined correspondence can be understood as a pre-established lookup table or model that associates different similarity ranges with corresponding water film thicknesses. Specifically, this correspondence can be obtained through experimental measurement, physical modeling, or machine learning. For example, multispectral images can be collected under water films of different thicknesses, and the similarity between their spectral curves and the spectral curves of dried leaves can be calculated to establish a mapping relationship between similarity and thickness. The similarity range refers to dividing the similarity values ​​into several intervals, each interval corresponding to a specific water film thickness. In this way, when a similarity value is calculated, the corresponding water film thickness can be directly found based on its corresponding similarity range.

[0115] This application provides a concrete and efficient method for determining the thickness of a water film by introducing a second preset correspondence. After obtaining the similarity between the spectral curve of the water film and a preset drying spectral curve, this similarity value is used to search within the pre-established second preset correspondence. Since this correspondence has already pre-associated the similarity range with the water film thickness, the corresponding water film thickness can be quickly determined directly based on the range of the similarity value. This lookup table or model-based determination method avoids complex real-time calculations and simplifies the thickness determination process.

[0116] The above technical solution provides a specific, convenient, and highly operable method for determining water film thickness. This method utilizes a pre-established second preset correspondence to directly map the calculated similarity value to the corresponding water film thickness, thereby improving the efficiency and accuracy of water film thickness determination and providing a reliable parameter basis for subsequently eliminating optical distortion introduced by liquid water.

[0117] This application also discloses an intelligent detection system for garden plant diseases and pests based on image recognition, comprising: an acquisition device and a processing device; the acquisition device is used to acquire multispectral images of garden plants; the processing device is used to determine whether liquid water exists on the leaf surface of the garden plants based on the reflectance of the multispectral images; the processing device is used to determine the type of liquid water in the multispectral images based on image features when liquid water exists on the leaf surface of the garden plants; the processing device is used to process the multispectral images based on the water type to eliminate the optical distortion introduced by the liquid water and obtain a corrected multispectral image; the processing device is used to judge plant diseases and pests based on the corrected multispectral images.

[0118] This system aims to address the problem in traditional intelligent detection of diseases and pests in garden plants. Due to the complexity of the garden environment, especially when liquid water (such as water droplets or films) is present on plant leaves, optical distortion caused by these water bodies during image acquisition makes it difficult for intelligent recognition systems to accurately distinguish between pathological changes and optical phenomena, leading to frequent false alarms. The system described above allows the acquisition device to collect multispectral images of garden plants, while the processing device intelligently identifies the presence and type of liquid water on the leaf surface. It then performs targeted image correction to eliminate optical distortion, ultimately making an accurate disease and pest diagnosis based on the corrected image. This collaborative mechanism ensures detection accuracy and reliability in complex environments, effectively enhancing the system's practical value.

[0119] The core of the intelligent detection system for garden plant diseases and pests based on image recognition proposed in this application lies in the collaborative work of the acquisition device and the processing device to identify and correct the optical distortion introduced by liquid water, and finally make an accurate judgment on diseases and pests.

[0120] Specifically, the acquisition device is configured to acquire multispectral images of garden plants. This acquisition device can be a multispectral camera integrated into a drone for large-scale, high-efficiency image acquisition; alternatively, it can be a handheld multispectral analyzer, allowing operators to photograph target plants at close range from the ground; or it can consist of a fixedly mounted array of multispectral sensors for continuous monitoring of garden plants in a specific area.

[0121] The processing device is configured to perform a series of image analysis and processing tasks. First, it determines whether liquid water is present on the leaf surface of garden plants based on the reflectance of multispectral images. Specifically, the processing device can be programmed to analyze the reflectance characteristics of different bands in the multispectral image; for example, by calculating the difference or ratio of reflectance in specific bands sensitive to water (such as near-infrared and green bands) and comparing it with a preset threshold to determine the presence of liquid water. Alternatively, the processing device can integrate a pre-trained machine learning model that learns reflectance patterns from a large number of multispectral image samples containing and not containing liquid water to determine the presence of liquid water in new images.

[0122] Furthermore, when liquid water is present on the surface of garden plant leaves, the processing device is used to determine the type of liquid water in the multispectral image based on image features. For example, the processing device can be configured to extract image features such as brightness values ​​and spectral curves of the liquid water region. By analyzing these features, such as water droplets typically appearing as bright, well-defined circular or elliptical areas, while water films may appear as areas with relatively uniform brightness and a larger coverage area, the processing device can distinguish between these two main types of water bodies: water droplets and water films.

[0123] Based on this, the processing device processes the multispectral image according to the determined water body type to eliminate the optical distortion introduced by liquid water, obtaining a corrected multispectral image. Specifically, the processing device can employ different correction strategies for different water body types. For example, when the water body type is water droplets, the processing device can invoke a preset optical model that can simulate the lens effect of water droplets and calculate a correction factor based on parameters such as the size and shape of the water droplets, thereby adjusting the pixel brightness of the affected area. When the water body type is a water film, the processing device can establish a reflectance correction model based on the thickness and optical characteristics of the water film, adjusting the spectral reflectance of the area covered by the water film to restore the true reflectance of the leaf.

[0124] Finally, the processing device is used to determine plant diseases and pests based on the corrected multispectral images. The processing device can integrate various image recognition algorithms; for example, it can utilize deep learning models, such as convolutional neural networks (CNNs), to analyze the corrected images and identify the unique image features of different diseases and pests. Furthermore, the processing device can also combine traditional image processing methods, such as texture analysis, color analysis, and shape analysis, to extract features of diseased areas and compare them with a pre-set disease feature database for disease classification and identification.

[0125] The intelligent detection system for garden plant diseases and pests based on image recognition proposed in this application effectively solves the problem of decreased detection accuracy caused by liquid water on leaf surfaces in complex garden environments by introducing a collaborative working mechanism between the acquisition and processing devices. Compared with the closest existing technology, the advantage of this application lies in its refined processing of environmental factors. Existing technologies often directly judge diseases and pests from multispectral images, ignoring the potential impact of liquid water on image quality, resulting in a high false alarm rate in humid environments such as after rain or with dew. This application effectively eliminates these interfering factors by adding liquid water detection and optical distortion correction steps before disease and pest judgment, enabling subsequent disease and pest judgment to be based on more realistic and accurate plant image information. Therefore, this application can significantly improve the accuracy and reliability of intelligent detection of garden plant diseases and pests, reduce the additional manpower and material resources consumed due to false alarms, and enhance users' trust in the intelligent detection system, thus having significant practical application value.

[0126] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent detection of diseases and pests in garden plants based on image recognition, characterized in that, include: Acquire multispectral images of garden plants; The presence of liquid water on the leaf surface of garden plants is determined based on the reflectance of the multispectral images. When the liquid water is present on the surface of the leaves of garden plants, the type of water in the liquid water in the multispectral image is determined based on the image features of the multispectral image; The multispectral image is processed based on the water body type to eliminate the optical distortion introduced by the liquid water body, resulting in a corrected multispectral image. Plant diseases and pests are identified based on the corrected multispectral images.

2. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 1, characterized in that, Determining the presence of liquid water on the leaf surface of garden plants based on the reflectance of the multispectral image includes: A first image in the near-infrared band and a second image in the green band are separated from the multispectral image; The presence of liquid water on the leaf surface of garden plants is determined based on the reflectance of the first image and the reflectance of the second image.

3. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 2, characterized in that, Determining whether liquid water exists on the surface of garden plant leaves based on the reflectance of the first image and the reflectance of the second image includes: The ratio of the first value and the second value is used as the normalized differential water index; the first value is the difference between the reflectance of the first image and the reflectance of the second image, and the second value is the sum of the reflectance of the first image and the reflectance of the second image. The presence of liquid water on the leaf surface of garden plants is determined based on the normalized differential water index and the brightness value of the second image.

4. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 3, characterized in that, Determining whether liquid water exists on the leaf surface of garden plants based on the normalized differential water index and the brightness value of the second image includes: Determine whether the normalized differential water index is less than a preset water index threshold; If the normalized differential water index is not less than the preset water index threshold, it is determined that there is no liquid water on the leaf surface of the garden plant. If the normalized differential water index is less than a preset water index threshold, determine whether the number of bright pixels in the second image is greater than a preset number threshold; the brightness value of the bright pixels is greater than a preset brightness threshold. If the number of highlighted pixels in the second image is greater than a preset threshold, it is determined that there is liquid water on the surface of the leaves of the garden plant; otherwise, it is determined that there is no liquid water on the surface of the leaves of the garden plant.

5. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 1, characterized in that, The image features include brightness values ​​and spectral curves. When liquid water is present on the surface of the leaves of garden plants, the water type of the liquid water in the multispectral image is determined based on the image features of the multispectral image, including: Based on the brightness values ​​of the multispectral image, regions in the multispectral image with brightness values ​​greater than a preset brightness threshold are identified as liquid water bodies. The water type of the liquid water body is determined based on its spectral curve.

6. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 5, characterized in that, Water body types include water droplets or water films. The water body type is determined based on the spectral curve of the liquid water body, including: Determine the similarity between the spectral curve of the liquid water and the preset drying spectral curve of the leaf surface in the multispectral image; When the similarity is less than a preset similarity threshold, the water body type of the liquid water body is determined to be a water film; otherwise, the water body type of the liquid water body is determined to be a water droplet.

7. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 5, characterized in that, When the water body type is water droplets, the multispectral image is processed based on the water body type to eliminate the optical distortion introduced by the liquid water body, resulting in a corrected multispectral image, including: Call the preset optical simulation model; Obtain the first brightness value of each pixel in the region corresponding to the water droplet; the first brightness value is the brightness value of each pixel in the region corresponding to the water droplet on the leaf of the garden plant, determined according to the preset optical simulation model when there is no water droplet on the leaf surface of the garden plant. Measure the diameter of the liquid water body in the multispectral image; The diameter of the liquid water body is input into the preset optical simulation model to obtain the second brightness value of each pixel in the region corresponding to the water droplet on the leaf of the garden plant; The difference between the second brightness value and the first brightness value is taken as the brightness value difference; The difference between the current brightness value and the corresponding brightness value of each pixel in the region corresponding to the water droplet on the leaf of the garden plant in the multispectral image is used as the corrected brightness value to obtain the corrected multispectral image.

8. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 5, characterized in that, When the water body type is a water film, the multispectral image is processed based on the water body type to eliminate the optical distortion introduced by the liquid water body, resulting in a corrected multispectral image, including: Obtain a first preset correspondence; the first preset correspondence includes a one-to-one correspondence between multiple thickness ranges and multiple adjustment coefficients; The thickness of the water film is determined based on the similarity between the spectral curve of the water film and the preset drying spectral curve of the leaf surface of the garden plant. The adjustment coefficient corresponding to the thickness range of the water film in the first preset correspondence is taken as the target adjustment coefficient; The corrected multispectral image is obtained by multiplying the current spectral reflectance of the region corresponding to the water film in the multispectral image with the target adjustment coefficient.

9. The intelligent detection method for garden plant diseases and pests based on image recognition according to claim 8, characterized in that, The thickness of the water film is determined based on the similarity between its spectral curve and a preset drying spectral curve of the leaf surface of the garden plant, including: Obtain a second preset correspondence; the second preset correspondence includes a one-to-one correspondence between multiple similarity ranges and multiple thicknesses; The thickness of the water film is defined as the thickness of the similarity range within the second preset correspondence.

10. An intelligent detection system for garden plant diseases and pests based on image recognition, characterized in that, include: Acquisition device and processing device; The acquisition device is used to acquire multispectral images of garden plants; The processing device is used to determine whether there is liquid water on the surface of the leaves of garden plants based on the reflectance of the multispectral image. The processing device is used to determine the type of liquid water in the multispectral image based on the image features of the multispectral image when the liquid water is present on the surface of the leaves of garden plants. The processing device is used to process the multispectral image based on the water body type to eliminate the optical distortion introduced by the liquid water body and obtain a corrected multispectral image. The processing device is used to determine plant diseases and pests based on the corrected multispectral images.

Images