Lawn coverage evaluation platform based on image recognition

Abstract

The invention discloses a lawn coverage evaluation platform based on image recognition, which belongs to the technical field of lawn measurement and comprises an image acquisition module, an image preprocessing module, a lawn area recognition module, a coverage calculation module, a scoring and analysis module and a result output module which are connected in sequence. The method comprises the following steps: performing standardized image acquisition on a lawn through an image acquisition module, and preprocessing an image to realize geometric standardization of the image; an artificial intelligence image segmentation model is adopted to improve the accuracy and adaptability of lawn area identification; the sample areas are scored and analyzed by calculating the coverage, and the method is used for comparative analysis of multiple varieties. By adopting the method, the acquisition efficiency, the identification accuracy and the evaluation standardization level of the coverage data in the lawn grass seed regional test can be improved, and the problems of non-uniform acquisition conditions, fragmentation of an identification method, incomparable results and the like in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of lawn measurement, and in particular relates to a lawn coverage assessment platform based on image recognition. Background Art

[0002] Regional trials of turfgrass varieties are a crucial step in turfgrass variety selection and evaluation. Turf cover (i.e., the proportion of ground surface covered by green turf tissue) is a key phenotypic indicator for evaluating turfgrass quality, growth potential, and stress resistance. Existing techniques for assessing turf cover primarily rely on manual scoring, visual estimation, or simple image estimation. These methods suffer from poor accuracy, strong subjectivity, and low reproducibility, making them difficult to meet the practical needs of regional trials for multiple varieties, large sample sizes, and standardized measurements.

[0003] In recent years, advances in artificial intelligence and image recognition technologies have opened new possibilities for obtaining plant phenotypic information. Some studies have attempted to use deep learning segmentation models to identify crop planting areas, but most are limited to natural images and lack standardized acquisition and analysis processes. In particular, for batch assessment of turfgrass cover, there is a lack of system integration solutions tailored to experimental scenarios. Therefore, developing an integrated platform for turfgrass cover identification in regional trials, combining standardized image capture with automated identification and analysis, has significant practical value and technological innovation. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a lawn coverage evaluation platform based on image recognition.

[0005] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0006] A lawn coverage evaluation platform based on image recognition includes an image acquisition module, an image preprocessing module, a lawn area recognition module, a coverage calculation module, a scoring and analysis module, and a result output module connected in sequence;

[0007] An image acquisition module is used to collect lawn images in the sample area;

[0008] Image preprocessing module, used to identify the scale position, calculate the perspective deformation matrix, and perform image perspective correction, color standardization, and brightness adjustment;

[0009] The lawn area recognition module is used to perform pixel-level segmentation on the lawn area in the image and output a grass / non-grass mask map;

[0010] Cover calculation module, used to count the pixel ratio of lawn area and output coverage;

[0011] The scoring and analysis module obtains the image color characteristics based on the lawn mask image and performs a comprehensive score on the sample area;

[0012] The result output module is used to display the recognition images, scoring results and export data.

[0013] Furthermore, the image acquisition module is set on a mobile phone or camera, and the specific operations are as follows:

[0014] The operator uses a mobile phone or camera to take an image from directly above the lawn sample square, and places standard rulers at the four corners of the image.

[0015] Furthermore, the preprocessing steps of the image preprocessing module are as follows:

[0016] (1) Perspective correction: Based on the coordinates of the four scale corner points in the image, the transformation matrix is ​​derived through perspective transformation to restore the quadrat image from the perspective view under the shooting angle to the top-down orthographic image;

[0017] (2) Color standardization: extract the RGB value of the color card area in the image and normalize the image color;

[0018] (3) Brightness and background suppression: Convert to HSV color space, remove low-saturation, high-brightness or extremely dark pixels, and enhance the contrast of lawn area recognition.

[0019] Furthermore, the steps for normalizing the image color are as follows:

[0020] Assume that the standard RGB mean value of the color card area in the image is: R a ,G a ,B a ;

[0021] The average RGB value of the current overall area of ​​the image is: R - , B - ;

[0022] Assume that the standard RGB value of the color card is (Rc, Gc, Bc), and the average value of the image color area is (R - , B - ), the normalized coefficient is:

[0023]

[0024] The adjusted value of each pixel is:

[0025] (R',G',B')=(k R ·R,k G ·G,k B B);

[0026] Among them, the parameters R, G, and B are the original red, green, and blue channel brightness values ​​of any pixel in the input image; Rc, Gc, and Bc are the ideal RGB reference values ​​of the colorimetric card under standard lighting conditions; R, B is the average R, G, and B channel values ​​of the color card area in the current image, which is automatically extracted from the image; R', G', and B' are the normalization coefficients of the R, G, and B channels respectively.

[0027] Furthermore, the lawn area recognition module uses a pixel-level semantic segmentation model based on U-Net to perform lawn recognition, and the steps are as follows:

[0028] (1) Training data preparation: Use manually labeled lawn image samples to construct a training set;

[0029] (2) Model training configuration: The loss function uses a combination of Binary Cross Entropy and Dice Loss;

[0030] (3) Segmentation output: After training, the input image is output with a mask image of the same size as the original image. The pixel value of the lawn area is 1 and the background is 0. The result is used for the lawn coverage calculation and quality scoring module.

[0031] Furthermore, the calculation formula of the coverage is as follows:

[0032] G=(Ng / Nt)×100%

[0033] Where G is coverage, %; Ng is the total number of pixels in the lawn area; Nt is the total number of pixels in the mask area.

[0034] Furthermore, the scoring and analysis module extracts color features of the lawn area in the image from the obtained lawn mask image, and the color features include green intensity, color saturation and mottle index.

[0035] Furthermore, the green intensity is the mean value of the pixels extracted from the G channel, the color saturation is the mean value of the S channel in the lawn area after converting the image to the HSV color space, and the mottle index is the pixel standard deviation calculated after converting the lawn area to a grayscale image. The comprehensive score S is calculated as follows:

[0036]

[0037] Among them, α, β, and γ are score weighting coefficients.

[0038] Furthermore, the result output module is connected to a display, which displays the original image, lawn mask map, coverage value and scoring result, and can perform statistical analysis and archiving of the test data of the sample area.

[0039] Furthermore, in the result table exported by the result output module, the fields include image number, acquisition time, sample number, coverage, green intensity, and scoring level, which are exported to CSV or Excel format.

[0040] Compared with the prior art, the present invention has the following technical advances:

[0041] This invention uses an image acquisition module to capture standardized lawn images and preprocesses them to achieve geometric standardization. It also employs an artificial intelligence image segmentation model to improve the accuracy and adaptability of lawn area recognition. Sample areas are scored and analyzed by calculating cover, enabling comparative analysis across multiple varieties. This invention improves the efficiency of acquiring cover data, identification accuracy, and evaluation standardization in regional lawn grass trials, resolving issues such as inconsistent acquisition conditions, fragmented identification methods, and incomparable results in existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0043] In the attached figure:

[0044] Figure 1 A schematic diagram of the process of a lawn coverage assessment platform based on image recognition provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. The embodiments of the present invention will be described below with reference to the accompanying drawings.

[0046] An embodiment of the present invention provides a lawn cover assessment platform based on image recognition, comprising an image acquisition module, an image preprocessing module, a lawn area recognition module, a cover calculation module, a scoring and analysis module, and a result output module, all connected in sequence. The image acquisition module is used to capture lawn images within a sample area. The image preprocessing module is used to identify the scale position, calculate the perspective deformation matrix, and perform image perspective correction, color normalization, and brightness adjustment. The lawn area recognition module is used to perform pixel-level segmentation of lawn areas in the image and output a grass / non-grass mask. The cover calculation module is used to calculate the percentage of lawn area pixels and output the cover. The scoring and analysis module obtains image color features based on the lawn mask and performs a comprehensive score on the sample area. The lawn quality of each sample plot is scored for comparative analysis between grass species area trials. The result output module is used to display the recognized images, scoring results, and export related data. This solution improves the accuracy and adaptability of lawn area recognition through a standardized image acquisition design and the use of an artificial intelligence image segmentation model. Cover calculation, scoring, and analysis facilitate batch comparisons between multiple varieties.

[0047] like Figure 1 As shown, the specific process of the present invention is as follows:

[0048] S100, capturing lawn images: The image acquisition module uses a handheld device, such as a mobile phone or a portable camera, to capture images. The specific operations are as follows:

[0049] The operator uses a mobile phone or portable camera to capture an image from directly above the lawn plot, ensuring the lens is essentially vertical. Standard scale rulers (e.g., 10cm black and white strips) are placed at the four corners of the image capture area for subsequent geometric correction, i.e., image perspective correction and size conversion to improve image scale consistency and angular standardization. The rulers are easy to place, require no fixed bracket, and can be flexibly moved between various plots, making them suitable for rapid image acquisition.

[0050] When collecting images, try to do so around noon or use fill-in lighting equipment to ensure uniform lighting and avoid shadow interference. The shooting resolution should be no less than 300dpi, and metadata such as shooting time and sample number should be recorded.

[0051] S200, Image Preprocessing: Identify the four corner scales in the captured image, automatically calculate the image perspective deformation matrix, and perform perspective correction to correct for image angle and scale distortion, achieving geometric standardization. Extract the RGB values ​​of the colorimetric chart area and perform color standardization on the entire image. Convert to HSV space, remove extremely dark, extremely bright, or low-saturation pixels, and enhance contrast in the lawn area, improving overall image quality and the robustness of subsequent segmentation.

[0052] The specific steps are as follows:

[0053] (1) Perspective correction:

[0054] In order to restore the quadrat image from a perspective view under a shooting angle to a top-down orthographic image, the present invention adopts a perspective correction method based on scale corner points.

[0055] Assume that the pixel coordinates of the four corner points of the ruler in the image are:

[0056] (x1,y1),(x2,y2),(x3,y3),(x4,y4)

[0057] The corresponding coordinates in the ideal correction plane are:

[0058] (x1′,y1′),(x2′,y2′),(x3′,y3′),(x4′,y4′)

[0059] Then the image perspective transformation can be expressed as the following homogeneous coordinate relationship:

[0060] [x′y′1]T=H·[xy1]T

[0061] The matrix H is a 3×3 perspective transformation matrix containing 8 independent variables (the 9th item is usually normalized to 1), which is used to describe the projection mapping from the original image plane to the target top-view plane.

[0062] When more than 4 pairs of corresponding points are known When , the following linear equations can be constructed:

[0063] A·h=b

[0064] Where A is the coefficient matrix, b is the target coordinate column vector, and h is the expanded column vector form of H. This system of equations can be solved by the least squares method (LeastSquares).

[0065] During image processing, the system uses the getPerspectiveTransform() function in the OpenCV library to calculate H and calls the warpPerspective() function to perform perspective correction, outputting a standardized overhead image. This image has uniform scale, angle, and distortion control characteristics, providing a consistent input basis for lawn area recognition.

[0066] (2) Color standardization: In order to eliminate the interference of light intensity, white balance setting, etc. on the color characteristics of lawn images under different shooting conditions, the present invention introduces a colorimetric card as a standard color reference and uses the grayscale world principle to normalize the image color.

[0067] Assume that the standard RGB mean value of the color card area in the image is: R a ,Ga ,B a

[0068] The average RGB value of the current overall area of ​​the image is: R - , B - ;

[0069] Assume that the standard RGB value of the color card is (Rc, Gc, Bc), and the average value of the image color area is (R - , B - ), the normalized coefficient is:

[0070]

[0071] The adjusted value of each pixel is:

[0072] (R',G',B')=(k R ·R,k G G, k B B)

[0073] Among them, the parameters R, G, and B are the original red, green, and blue channel brightness values ​​of any pixel in the input image; Rc, Gc, and Bc are the ideal RGB reference values ​​of the colorimetric card under standard lighting conditions; R, B - The average R, G, and B channel values ​​of the color card area in the current image are automatically extracted from the image; R', G', and B' are the normalization coefficients of the R, G, and B channels respectively.

[0074] The above normalization operation can uniformly adjust the color channels of the entire image, making the image color closer to the standard reference state, thereby improving the consistency and reliability of subsequent lawn area identification and scoring.

[0075] (3) Brightness and background suppression: Convert to HSV color space, remove low-saturation, high-brightness or extremely dark pixels, and enhance the contrast of lawn area recognition.

[0076] 1) In the image preprocessing module, the original lawn image is first converted from BGR color space to HSV color space. This conversion is performed using the OpenCV function cv2.cvtColor(Img_BGR,cv2.COLOR_BGR2HSV), which takes the original image Img_BGR as input and outputs an HSV image containing three channels: H (hue), S (saturation), and V (brightness). This conversion effectively separates color information from illumination information, providing a stable color representation for subsequent lawn pixel extraction and background removal.

[0077] 2) In the image preprocessing module, the system sets threshold ranges for hue (H), saturation (S), and brightness (V) in the HSV color space, for example, H∈[30,90], S≥40, and V∈[50,230]. The inRange() function in OpenCV is used to construct a binary mask to remove anomalous pixels in non-grass areas. A pixel value of 255 in the mask indicates that the pixel is within the set threshold range and is retained for subsequent lawn area identification. The bitwise_and() function is then called to apply the mask to the original image, achieving color enhancement and background suppression in the lawn area, improving the recognition accuracy and robustness of the subsequent segmentation model.

[0078] S300, Lawn Identification: Use a U-Net-based pixel-level semantic segmentation model to perform pixel-level identification of lawn areas in the image and output a grass / non-grass binary mask. The specific steps are as follows:

[0079] (1) Training data preparation: A training set is constructed using manually annotated lawn image samples. The mask image uses 1 to represent lawn pixels and 0 to represent background pixels. The image size is uniformly set to 256×256 or 512×512. Enhancement methods include rotation, flipping, brightness perturbation, etc. The training samples cover different grass species, different lighting conditions, and background environments.

[0080] (2) Model training configuration:

[0081] Loss function: Binary Cross Entropy + Dice Loss combination;

[0082] -Optimizer: Adam, initial learning rate 1e-4;

[0083] -Batch size: 8-16; training epochs: 50-100 epochs;

[0084] -Use the validation set to evaluate IoU and accuracy, and save the best model for inference.

[0085] (3) Segmentation output: The trained U-Net-based pixel-level semantic segmentation model outputs a mask image of the same size as the original image, with the pixel value of the lawn area being 1 and the background being 0. The result is used in the lawn coverage calculation and quality scoring module.

[0086] The above-mentioned pixel-level semantic segmentation model based on U-Net takes the input image as feature input and outputs a grass / non-grass pixel probability map. And compared with the manually annotated label map Y∈{0,1} H×W , calculate the binary cross entropy loss as follows:

[0087]

[0088] Where, represents the average loss value of the entire image, which is used as the objective function for model training; H represents the image height (pixels); W represents the image width (pixels); Yij represents the true label value of position (i, j), where lawn is 1 and non-lawn is 0; Indicates the probability value predicted by the model at position (i, j) as lawn, ranging from [0,1];

[0089] After the above model training is completed, the probability map is binarized by setting a threshold and the lawn mask map is output.

[0090] S400, coverage calculation:

[0091] To achieve automated and standardized lawn coverage assessment, the present invention uses pixel statistics to calculate lawn coverage based on image mask results. The specific formula is derived as follows:

[0092] After the image is processed by the deep learning segmentation model, a grass / non-grass binary classification mask is generated. Let the total number of pixels in the mask be: Nt = W × H;

[0093] Where W is the image width (pixels); H is the image height (pixels);

[0094] Assuming that the lawn area pixels are represented by a value of 1 in the mask image and the background is 0, the total number of lawn pixels Ng is expressed as follows:

[0095]

[0096] Where M(i, j) is the value of the i-th and j-th pixel in the mask image, 1 represents grass and 0 represents non-grass;

[0097] The calculation formula of lawn coverage G is as follows:

[0098]

[0099] Where G is coverage, %; Ng is the total number of pixels in the lawn area; Nt is the total number of pixels in the mask area.

[0100] S500, Turfgrass Quality Scoring and Analysis: Combined with coverage data and image color information, the scoring and analysis module evaluates the turfgrass and assigns a turfgrass quality score to each sample plot. This can be used for comparative analysis between grass seed area trials. The specific steps are as follows:

[0101] (1) Input basis:

[0102] 1) Use the mask image output by the lawn area recognition module to extract lawn area pixels;

[0103] 2) Based on the mask map, the original image is masked and only the lawn area is retained for color feature calculation.

[0104] (2) Calculation indicators:

[0105] 1) Green intensity: extract the pixel mean of the G channel to reflect the greenness level of the lawn;

[0106] 2) Color saturation: Convert the image to the HSV color space and extract the mean value of the S channel in the lawn area to measure the vividness of the color;

[0107] 3) Mottle index: After converting the lawn area into a grayscale image, the standard deviation of its pixels is calculated to reflect the consistency of the lawn color. The higher the value, the more obvious the mottle phenomenon.

[0108] Calculate the overall score:

[0109]

[0110] Among them, α, β, and γ are scoring weighting coefficients. The present invention is trained on 200 lawn sample images. Through iterative calculation of loss function minimization, the weighting coefficients α=0.4, β=0.4, and γ=0.2 are finally determined, which have strong stability.

[0111] S600, result output (data archiving and display):

[0112] (1) Real-time display: The original image, lawn mask, coverage value and scoring results are displayed on the display;

[0113] -Batch export: automatically generate structured data tables (Excel or CSV format);

[0114] -Data archiving: Categorized and stored by sample number, shooting time and other information to support long-term review and retrospective analysis.

[0115] (2) Output data results:

[0116] In the system export result table, the fields include:

[0117] -Image ID

[0118] -Plot ID

[0119] -Capture_Time

[0120] - Lawn coverage (Cover_%)

[0121] - Green intensity (G_mean)

[0122] -Color saturation (S_mean)

[0123] - Texture_std

[0124] -Comprehensive score (Score_S)

[0125] -Grade

[0126] (3) Display interface function (can be deployed on a web page or local application):

[0127] - Image display: Display the original image, mask image and score value in parallel with pictures and text;

[0128] -Data table browsing: Browse the rating results in pages, and filter by grass species, time, rating level, etc.

[0129] - Download function: export the current screening results with one click;

[0130] -Batch management: supports archiving and switching browsing of multiple rounds of test data.

[0131] In summary, the present invention has the advantages of simple operation and strong intuitiveness. It standardizes the image acquisition design, unifies the image acquisition scale and lighting, and reduces environmental interference; introduces a ruler and a colorimetric card to provide an image ratio and color calibration benchmark to enhance data consistency; adopts an artificial intelligence image segmentation model to improve the accuracy and adaptability of lawn area identification; constructs an automatic coverage calculation and scoring model to support batch comparisons between multiple varieties in lawn grass area tests; the platform displays the results through a display visualization interface for easy observation. The present invention can be widely used in scenarios such as lawn variety regional tests, lawn growth status monitoring, lawn stress resistance assessment, and lawn construction quality inspection, and has significant promotion and application value. The integrated platform provided by the present invention is easy to operate and has intuitive output, and is suitable for use by scientific research units, breeding institutions, lawn construction management and other departments.

[0132] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A lawn coverage assessment platform based on image recognition, characterized by: It includes an image acquisition module, an image preprocessing module, a lawn area recognition module, a coverage calculation module, a scoring and analysis module and a result output module, which are connected in sequence; An image acquisition module is used to collect lawn images in the sample area; Image preprocessing module, used to identify the scale position, calculate the perspective deformation matrix, and perform image perspective correction, color standardization, and brightness adjustment; The lawn area recognition module is used to perform pixel-level segmentation on the lawn area in the image and output a grass / non-grass mask map; Cover calculation module, used to count the pixel ratio of lawn area and output coverage; The scoring and analysis module obtains the image color characteristics based on the lawn mask image and performs a comprehensive score on the sample area; The result output module is used to display the recognition images, scoring results and export data.

2. The lawn coverage assessment platform based on image recognition according to claim 1, characterized in that: The image acquisition module is set on a mobile phone or camera, and the specific operations are as follows: The operator uses a mobile phone or camera to take an image from directly above the lawn sample square, and places standard rulers at the four corners of the image.

3. The lawn coverage assessment platform based on image recognition according to claim 2, characterized in that: The preprocessing steps of the image preprocessing module are as follows: (1) Perspective correction: Based on the coordinates of the four scale corner points in the image, the transformation matrix is ​​derived through perspective transformation to restore the quadrat image from the perspective view under the shooting angle to the top-down orthographic image; (2) Color standardization: extract the RGB value of the color card area in the image and normalize the image color; (3) Brightness and background suppression: Convert to HSV color space, remove low-saturation, high-brightness or extremely dark pixels, and enhance the contrast of lawn area recognition.

4. The lawn coverage assessment platform based on image recognition according to claim 3, characterized in that: The steps to normalize the image colors are as follows: Assume that the standard RGB mean value of the color card area in the image is: R a ,G a ,B a ; The average RGB value of the current overall area of ​​the image is: R - , B - ; Assume that the standard RGB value of the color card is (Rc, Gc, Bc), and the average value of the image color area is The normalization coefficient is: The adjusted value of each pixel is: (R′,G′,B′)=(k R ·R,k G ·G,k B ·B); Among them, the parameters R, G, and B are the original red, green, and blue channel brightness values ​​of any pixel in the input image; R', G', and B' are the normalization coefficients of the R, G, and B channels respectively.

5. The lawn coverage assessment platform based on image recognition according to claim 4, characterized in that: The lawn area recognition module uses a pixel-level semantic segmentation model based on U-Net to perform lawn recognition. The steps are as follows: (1) Training data preparation: Use manually labeled lawn image samples to construct a training set; (2) Model training configuration: The loss function uses a combination of Binary Cross Entropy and Dice Loss; (3) Segmentation output: After training, a mask image of the same size as the original image is output, with the pixel value of the lawn area being 1 and the background being 0. The result is used in the lawn coverage calculation and quality scoring module.

6. The lawn coverage assessment platform based on image recognition according to claim 5, characterized in that: The calculation formula of the coverage is as follows: G=(Ng / Nt)×100% Where G is coverage, %; Ng is the total number of pixels in the lawn area; Nt is the total number of pixels in the mask area.

7. The lawn coverage assessment platform based on image recognition according to claim 65, characterized in that: The scoring and analysis module extracts color features of the lawn area in the image from the obtained lawn mask image, and the color features include green intensity, color saturation and mottle index.

8. The lawn coverage assessment platform based on image recognition according to claim 7, characterized in that: The green intensity is the mean value of the pixels extracted from the G channel, the color saturation is the mean value of the S channel in the lawn area after converting the image to the HSV color space, and the mottle index is the pixel standard deviation calculated after converting the lawn area to a grayscale image. The comprehensive score S is calculated as follows: Among them, α, β, and γ are score weighting coefficients.

9. The lawn coverage assessment platform based on image recognition according to any one of claims 1 to 8, characterized in that: The result output module is connected to a display, and displays the original image, lawn mask map, coverage value and scoring result on the display, which can perform statistical analysis and archiving of the test data of the sample area.

10. The lawn coverage assessment platform based on image recognition according to claim 9, characterized in that: In the result table exported by the result output module, the fields include image number, acquisition time, sample number, coverage, green intensity, and scoring level.

Images