A rice seed screening method based on data fusion

Abstract

This invention discloses a rice seed screening method based on data fusion, relating to the field of image processing technology, including the following steps: S1, acquiring sample images of rice seeds and extracting the RGB values ​​of pixels in the sample images; S2, processing using a neural network to determine the label value of each pixel in the sample image; S3, determining the location of diseased and pest-infested seeds. This invention integrates multi-dimensional image features, neural network pixel processing, and cluster analysis. First, it analyzes the differences in seed color distribution, utilizing the non-linear learning capability of the neural network combined with pixel-level label decision-making to achieve deep learning of color moment features; then, it quantifies the deviation of pixels within clusters through clustering processing, ultimately screening for diseased and pest-infested seeds and accurately marking the spatial location of diseases and pests. This provides quantitative and traceable technical support for rice seed quality control and can be directly applied to the breeding and quality testing of improved varieties in the seed industry.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a rice seed screening method based on data fusion. Background Technology

[0002] Disease and pest screening of rice seeds is a crucial step in ensuring rice planting quality and increasing yield. Currently, the industry mainly uses the following methods for seed screening:

[0003] Manual visual screening relies on professionals to visually inspect seeds for characteristics such as color and shape to determine the presence of pests and diseases. This method has significant drawbacks: extremely low efficiency, screening only a few hundred seeds per hour, which cannot meet the needs of large-scale seed production; unstable accuracy, affected by factors such as personnel experience and visual fatigue, with false negative and false positive rates often exceeding 15%; and high labor costs, as long-term reliance on a large number of professionals increases the operating costs of the seed industry.

[0004] Traditional machine vision screening: Some technologies attempt to assist screening with machine vision, but there are still shortcomings: feature extraction is singular, mostly relying only on raw RGB values ​​or simple color features, which cannot fully characterize the differences between normal seeds and diseased seeds; the classification model is simple and lacks the ability to learn high-dimensional nonlinear features, resulting in insufficient classification accuracy; there is a lack of clustering optimization for pixel-level labels and fine anomaly detection, making it difficult to accurately locate diseased areas, and the problems of missed detection and false detection are still prominent.

[0005] In summary, existing technologies have significant limitations in terms of efficiency, accuracy, and automation in screening for diseases and pests in rice seeds. There is an urgent need for a technical solution that can integrate the advantages of multiple technologies and achieve efficient and accurate screening. Summary of the Invention

[0006] To address the above problems, this invention proposes a rice seed screening method based on data fusion.

[0007] The technical solution of the present invention is: a rice seed screening method based on data fusion, comprising the following steps:

[0008] S1. Collect sample images of rice seeds and extract the RGB values ​​of pixels in the sample images;

[0009] S2. Based on the RGB values ​​of each pixel, a neural network is used to process the data and determine the label value of each pixel in the sample image.

[0010] S3. Cluster all pixels in the sample image according to their label values, and determine the location of the disease and pest seeds based on the clustering results.

[0011] Furthermore, S2 includes the following sub-steps:

[0012] S21. Construct the initial feature values ​​of the pixels based on their RGB values;

[0013] S22. Normalize the initial feature values ​​of the pixels to obtain standard feature values;

[0014] S23. Determine the label value of the pixel in the sample image based on the standard feature values ​​of the pixel.

[0015] The beneficial effect of the above-mentioned further solution is that, in this invention, the differences in the magnitude and distribution of the original RGB values ​​can interfere with neural network training, and direct use lacks a characterization of the color distribution pattern. Therefore, this invention enhances the discriminative power of the features by constructing initial features using color moments, eliminating differences through normalization, and then classifying them using a neural network.

[0016] Furthermore, in S21, the initial feature quantities of the sample image The expression is:

[0017] ;

[0018] in, The first moment of the R color of the pixel. Let R be the second moment of the color of the pixel. The third moment of the R color of a pixel, Let G be the first moment of the color of the pixel. Let G be the second moment of the color of the pixel. The third moment of the G color of a pixel. Let B be the first moment of the color of the pixel. Let B be the second moment of the color of the pixel. Let B be the third moment of the color of the pixel.

[0019] The beneficial effects of the above-mentioned further scheme are as follows: In this invention, the first moment characterizes the central tendency of color, the second moment characterizes the dispersion of color, and the third moment characterizes the symmetry of color distribution. Normal rice seeds and diseased seeds show significant differences in the statistical distribution of color (e.g., uneven color intensity and skewed distribution in moldy areas), and these differences cannot be captured using only the original RGB values. By combining three channels × the third moment, a 9-dimensional color moment is obtained.

[0020] Furthermore, S23 includes the following sub-steps:

[0021] S231. Input the standard feature values ​​of each pixel into the neural network, process them through the hidden layer of the neural network, and obtain the predicted label of the pixel.

[0022] S232. Determine the label value of the pixel in the sample image based on each element of the predicted label.

[0023] The beneficial effect of the above-mentioned further scheme is that, in this invention, the input standard feature quantity is subjected to multi-layer nonlinear transformation through the weight matrix, bias vector and activation function of the hidden layer, and the pixel-level predicted label is output.

[0024] Furthermore, in S231, the predicted label of the pixel. The expression is:

[0025] ;

[0026] in, These are the standard feature values ​​of a pixel. Let be the hidden layer bias vector of the neural network. This represents the hidden layer weight vector of the neural network. This represents the activation function.

[0027] Furthermore, in S232, the label value of the pixel in the sample image The expression is:

[0028] ;

[0029] in, The predicted label for each pixel contains the first... A vector, For vector modulo operation, The number of vectors contained in the predicted label of a pixel. This indicates that the label value corresponding to the maximum value is taken.

[0030] The beneficial effect of the above further solution is as follows: In this invention, the predicted label output by the neural network is a set containing I vectors, where the magnitude of each vector reflects its strength. The vector with the largest magnitude is selected as the final label for the pixel in the sample image. The neural network may output multiple candidate labels (such as "suspected normal" or "suspected pests"), leading to decision ambiguity. The larger the magnitude, the higher the confidence level of the label. Therefore, selecting the vector with the largest magnitude as the final label can specifically solve the decision ambiguity problem caused by multiple outputs.

[0031] Furthermore, S3 includes the following sub-steps:

[0032] S31. Based on the label values ​​of all pixels in the sample image, perform clustering to determine the cluster to which each pixel belongs;

[0033] S32. Among the pixels contained in each cluster, the largest RGB value is taken as the first standard value of the cluster, the smallest RGB value is taken as the second standard value, and the clustering feature value of each pixel in its respective cluster is calculated.

[0034] S33. In each cluster, pixels whose cluster feature value is less than half the mean of all cluster feature values ​​in that cluster are designated as pest and disease seeds.

[0035] The beneficial effects of the above further scheme are as follows: In this invention, for each cluster, the maximum RGB value (brightest / most vibrant color within the cluster) and the minimum RGB value (RGB_2) (darkest / lightest color within the cluster) are statistically analyzed; pixel color differences are calculated using Euclidean distance, and then cluster feature values ​​are obtained through exponential operations and fractional normalization, quantifying the RGB distribution position of pixels within the cluster. Pixel labels may contain noise (such as edge errors in neural network classification), and direct screening of pests and diseases based on labels is easily interfered with. By clustering semantically similar pixels first, and then performing anomaly detection on the RGB within each cluster, the problems of label noise and local feature differences can be specifically solved, improving the robustness of screening. In S33, the weight coefficient of the mean can be determined according to the actual situation; here, only 0.5 is taken as a representative value.

[0036] Furthermore, in S32, the clustering feature value of the pixel in its respective cluster is... The expression is:

[0037] ;

[0038] in, Represents the RGB value of a pixel. Indicates the first standard value. Indicates the second standard value. This indicates the calculation of Euclidean distance. This indicates exponent calculation.

[0039] The beneficial effects of this invention are as follows: This invention integrates multi-dimensional features of images, neural network pixel processing, and cluster analysis; firstly, it analyzes the differences in seed color distribution, and uses the non-linear learning ability of neural networks combined with pixel-level label decision-making to achieve deep learning of color moment features; then, it quantifies the degree of deviation of pixels within clusters through cluster processing, and finally screens seeds with pests and diseases, accurately marks the spatial location of pests and diseases, and provides quantitative and traceable technical support for the quality control of rice seeds, which can be directly applied to the breeding and quality testing of improved varieties in the seed industry. Attached Figure Description

[0040] Figure 1 This is a flowchart of a rice seed screening method based on data fusion. Detailed Implementation

[0041] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0042] like Figure 1As shown, this invention provides a rice seed screening method based on data fusion, comprising the following steps:

[0043] S1. Collect sample images of rice seeds and extract the RGB values ​​of pixels in the sample images;

[0044] S2. Based on the RGB values ​​of each pixel, a neural network is used to process the data and determine the label value of each pixel in the sample image.

[0045] S3. Cluster all pixels in the sample image according to their label values, and determine the location of the disease and pest seeds based on the clustering results.

[0046] In this embodiment of the invention, S2 includes the following sub-steps:

[0047] S21. Construct the initial feature values ​​of the pixels based on their RGB values;

[0048] S22. Normalize the initial feature values ​​of the pixels to obtain standard feature values;

[0049] S23. Determine the label value of the pixel in the sample image based on the standard feature values ​​of the pixel.

[0050] In this invention, the differences in the magnitude and distribution of the original RGB values ​​can interfere with neural network training, and direct use lacks a characterization of color distribution patterns. Therefore, this invention enhances the discriminative power of the features by constructing initial features using color moments, eliminating differences through normalization, and then classifying them using a neural network.

[0051] In this embodiment of the invention, in S21, the initial feature quantity of the sample image The expression is:

[0052] ;

[0053] in, The first moment of the R color of the pixel. Let R be the second moment of the color of the pixel. The third moment of the R color of a pixel, Let G be the first moment of the color of the pixel. Let G be the second moment of the color of the pixel. The third moment of the G color of a pixel. Let B be the first moment of the color of the pixel. Let B be the second moment of the color of the pixel. Let B be the third moment of the color of the pixel.

[0054] In this invention, the first moment characterizes the central tendency of color, the second moment characterizes the dispersion of color, and the third moment characterizes the symmetry of color distribution. Normal rice seeds and diseased seeds show significant differences in the statistical distribution of color (e.g., uneven color intensity and skewed distribution in moldy areas), which cannot be captured using only the original RGB values. By combining three channels × the third moment, a 9-dimensional color moment is obtained.

[0055] In this embodiment of the invention, S23 includes the following sub-steps:

[0056] S231. Input the standard feature values ​​of each pixel into the neural network, process them through the hidden layer of the neural network, and obtain the predicted label of the pixel.

[0057] S232. Determine the label value of the pixel in the sample image based on each element of the predicted label.

[0058] In this invention, the input standard features are subjected to multi-layer nonlinear transformation through the weight matrix, bias vector and activation function of the hidden layer, and the pixel-level predicted label is output.

[0059] In this embodiment of the invention, in S231, the predicted label of the pixel is... The expression is:

[0060] ;

[0061] in, These are the standard feature values ​​of a pixel. Let be the hidden layer bias vector of the neural network. This represents the hidden layer weight vector of the neural network. This represents the activation function.

[0062] In this embodiment of the invention, in S232, the label value of the pixel in the sample image The expression is:

[0063] ;

[0064] in, The predicted label for each pixel contains the first... A vector, For vector modulo operation, The number of vectors contained in the predicted label of a pixel. This indicates that the label value corresponding to the maximum value is taken.

[0065] In this invention, the predicted label output by the neural network is a set of I vectors, where the magnitude of each vector reflects its strength. The vector with the largest magnitude is selected as the final label for the pixel in the sample image. The neural network may output multiple candidate labels (such as "suspected normal" or "suspected pests"), leading to decision ambiguity. A larger magnitude indicates a higher confidence level for the label; therefore, selecting the vector with the largest magnitude as the final label can specifically address the decision ambiguity problem caused by multiple outputs.

[0066] In this embodiment of the invention, S3 includes the following sub-steps:

[0067] S31. Based on the label values ​​of all pixels in the sample image, perform clustering to determine the cluster to which each pixel belongs;

[0068] S32. Among the pixels contained in each cluster, the largest RGB value is taken as the first standard value of the cluster, the smallest RGB value is taken as the second standard value, and the clustering feature value of each pixel in its respective cluster is calculated.

[0069] S33. In each cluster, pixels whose cluster feature value is less than half the mean of all cluster feature values ​​in that cluster are designated as pest and disease seeds.

[0070] In this invention, for each cluster, the maximum RGB value (brightest / most vibrant color within the cluster) and the minimum RGB value (RGB_2) (darkest / lightest color within the cluster) are statistically analyzed. Pixel color differences are calculated using Euclidean distance, and then cluster feature values ​​are obtained through exponential operations and fractional normalization, quantifying the RGB distribution position of pixels within the cluster. Pixel labels may contain noise (such as edge errors in neural network classification), making direct label-based pest and disease screening susceptible to interference. By clustering semantically similar pixels and then performing anomaly detection on the RGB values ​​within each cluster, the problems of label noise and local feature differences can be specifically addressed, improving the robustness of the screening process. In S33, the weighting coefficient of the mean can be determined according to the actual situation; here, only 0.5 is used as a representative value.

[0071] In this embodiment of the invention, in S32, the pixel point is in the clustering feature value of its respective cluster. The expression is:

[0072] ;

[0073] in, Represents the RGB value of a pixel. Indicates the first standard value. Indicates the second standard value. This indicates the calculation of Euclidean distance. This indicates exponent calculation.

[0074] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for screening rice seeds based on data fusion, characterized in that, The method comprises the following steps: S1, collecting a sample image of rice seeds and extracting the RGB value of each pixel point in the sample image; S2, processing the RGB value of each pixel point by using a neural network to determine the label value of each pixel point in the sample image; S3, clustering the label value of all pixel points in the sample image, and determining the position of the diseased and insect-infested seeds according to the clustering result; The S3 comprises the following sub-steps: S31, clustering all the label values of the pixel points in the sample image to determine the cluster to which each pixel point belongs; S32, in each cluster, taking the maximum RGB value as the first standard value and the minimum RGB value as the second standard value, and calculating the cluster characteristic value of each pixel point in the cluster; S33, in each cluster, taking the pixel point whose cluster characteristic value is less than half of the average of all cluster characteristic values of the cluster as the diseased and insect-infested seed; In the S32, the pixel point is in the cluster characteristic value of the cluster to which the pixel point belongs The expression is: ； wherein, represents the RGB value of a pixel point, represents a first standard value, represents a second standard value, represents a Euclidean distance calculation, represents an exponential calculation.

2. The data fusion based paddy seed screening method according to claim 1, wherein, The S2 comprises the following sub-steps: S21, constructing the initial feature quantity of the pixel point according to the RGB value of the pixel point; S22, normalizing the initial feature quantity of the pixel point to obtain the standard feature quantity; S23, determining the label value of the pixel point in the sample image according to the standard feature quantity of the pixel point.

3. The data fusion based paddy seed screening method according to claim 2, wherein, In the S21, the initial feature amount of the sample image The expression is: ； wherein, is the R color first moment of the pixel point, is the R color second moment of the pixel point, is the R color third moment of the pixel point, is the G color first moment of the pixel point, is the G color second moment of the pixel point, is the G color third moment of the pixel point, is the B color first moment of the pixel point, is the B color second moment of the pixel point, is the B color third moment of the pixel point.

4. The data fusion based paddy seed screening method according to claim 2, wherein, The S23 comprises the following sub-steps: S231, inputting the standard feature quantity of each pixel point into the neural network and processing it through the hidden layer of the neural network to obtain the predicted label of the pixel point; S232, determining the label value of the pixel point in the sample image according to each element of the predicted label.

5. The data fusion based paddy seed screening method according to claim 4, wherein, In the S231, the prediction label of the pixel point The expression is: ； wherein, is a standard feature quantity of a pixel point, is a hidden layer bias vector of a neural network, is a hidden layer weight vector of a neural network, denotes an activation function.

6. The data fusion based paddy seed screening method according to claim 4, wherein, In the S232, the label value of the pixel point in the sample image The expression is: ； wherein, is a prediction label of a pixel point, is a vector contained in the prediction label of the pixel point, is a vector modulo operation, is a number of vectors contained in the prediction label of the pixel point, denotes a label value corresponding to the maximum value.

Images