Evaluation method of wheat fusarium wilt occurrence in the field based on improved YoloV4

Through improved YoloV4 target detection network and random forest classification, combined with Resnet network to extract the diseased spike characteristics, the problem of limited image acquisition conditions and low detection accuracy in the assessment of the incidence of wheat gibberellia is solved, and higher evaluation accuracy is achieved and faster model training time, suitable for assessment of the incidence of wheat gibberellia in complex conditions.

CN114120203BActive Publication Date: 2024-11-26ANHUI UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111452227.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-01
Publication Date
2024-11-26
Estimated Expiration
2041-12-01

AI Technical Summary

Technical Problem

The existing method of evaluating the incidence of wheat gibberellia is due to the limited image acquisition conditions, low accuracy of wheat ear detection and insignificant distinction between diseased ear characteristics, which leads to the inability to effectively promote and apply.

Method used

The improved YoloV4 object detection network model combined with random forest classification was used to adaptively segment wheat ear images and extract disease ear characteristics using Resnet network to calculate the disease ear rate to evaluate the degree of gibberellosis.

Benefits of technology

It improves the accuracy of wheat ear counting, shortens the model training time, enhances the accuracy and generality of evaluation, and can quickly and effectively evaluate the incidence of wheat gibberellia under complex conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114120203B_ABST
    Figure CN114120203B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for evaluating the occurrence degree of wheat scab in the field based on improved YoloV4, comprising the following steps: S100, shooting a group wheat ear image in an observation area; S200, using a trained improved YoloV4 target detection network model to detect the group wheat ear image in the observation area and obtain a single wheat ear image; S300, counting the number of wheat ears to obtain the total number of wheat ears, and performing feature extraction and random forest classification on the single wheat ear image and counting the number of diseased wheat ears; S400, calculating the diseased ear rate of the group wheat ears according to the total number of wheat ears and the number of diseased wheat ears, and obtaining the occurrence degree of scab of the group wheat ears according to relevant national standards. The method improves the prediction accuracy of the occurrence degree of wheat scab, can be used under complex conditions, does not require auxiliary equipment or materials, has good versatility, effectively solves the situation that the wheat ears are blocked and cannot be detected due to the shooting angle, the detection accuracy of the target detection network model is also improved, and the model training time is shortened.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of agricultural information technology, and in particular to a method for evaluating the occurrence degree of wheat fusarium head blight in the field based on improved YoloV4. Background Art

[0002] Wheat Fusarium Head Blight is a common wheat disease, and the rapid assessment of its occurrence is often used as an important basis for judging the severity of the disease in the field. The traditional statistical method is manual statistics, which requires plant protection personnel to visually check and count, which leads to a large amount of manpower and financial resources. With the popularization of deep learning technology in the agricultural field, the rapid assessment of the occurrence of wheat fusarium head blight will be greatly improved. For example, Wang et al. proposed a wheat fusarium head blight grade assessment method based on K-Means and CNN network in the article "Evaluation of Efficacy of Fungicides for Control of Wheat Fusarium Head Blight Based on Digital Imaging". This method is to roughly segment the wheat ear image through K-Means and enhanced green image (RGG), and then extract the color features and texture features of the segmented single wheat ear to identify the diseased condition of the wheat ear. Finally, the CNN network is used to count the wheat ears and calculate the diseased ear rate to obtain the occurrence of wheat fusarium head blight. Su et al. proposed a method based on dual Mask-RCNN networks to complete the evaluation of wheat head blight grade in the paper "Automatic evaluation of wheat resistance tofusarium head blight using dual mask-RCNN deep learning frameworks in computer vision". This method uses two Mask-RCNN networks to segment the group wheat ear image. The first Mask-RCNN network segments the wheat ears in the image and obtains and counts the individual wheat ear images. The obtained image is then segmented by the second Mask-RCNN network to obtain the diseased area of ​​the wheat ears and calculate the number of diseased wheat. Then, the group diseased ear rate is calculated and the disease grade of the wheat ears is determined.

[0003] The above technology has the following shortcomings: When sampling wheat pictures, the method proposed by Wang et al. needs to use black cloth to block the background of wheat to facilitate data processing. However, due to site restrictions or collection conditions, it is often impossible to use black cloth to block the background of wheat, which makes this method unable to be effectively promoted. At the same time, the counting accuracy of this method for wheat is low, resulting in the inability to obtain the accurate number of wheat ears, which reduces the prediction accuracy of the method. The method proposed by Su et al. is based on a dual Mask-RCNN network structure that is too complex, has problems such as long model training time and large amount of data annotation, and its promotion and application are also greatly limited. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In order to solve the technical problems that the image acquisition conditions for assessing the occurrence degree of wheat fusarium scabiei are limited, the wheat ear detection rate is low, and the characteristics of diseased ears are not clearly distinguished, the present invention proposes a method for assessing the occurrence degree of wheat fusarium scabiei in the field based on the improved YoloV4.

[0006] (II) Technical solution

[0007] To achieve the above objectives, the technical solution adopted by the present invention is: a method for evaluating the occurrence degree of wheat fusarium scabiei in the field based on improved YoloV4, comprising the following steps: S100, shooting images of group wheat ears in the observation area; S200, using the trained improved YoloV4 target detection network model to detect the group wheat ear images in the observation area and obtain single wheat ear images; S300, counting the number of wheat ears to obtain the total number of wheat ears; at the same time, feature extraction and random forest classification statistics are performed on the single wheat ear images to obtain the number of diseased wheat ears; S400, calculating the diseased ear rate of the group wheat ears according to the total number of wheat ears and the number of diseased wheat ears, and obtaining the occurrence degree of fusarium scabiei in the group wheat ears according to relevant national standards.

[0008] (III) Beneficial effects

[0009] Compared with the prior art, the present invention has the following positive technical effects. Compared with the existing method for evaluating the occurrence degree of wheat fusarium head blight, the present invention has a better counting accuracy of wheat ears, which will be beneficial to improving the prediction accuracy of the occurrence degree of the model. The present invention shortens the training time of the model a lot, can be put into use quickly, can be used under complex conditions, has good versatility, and the evaluation of the occurrence degree of wheat fusarium head blight is no longer limited by image acquisition conditions. The present invention also proposes an effective solution to the occlusion of wheat ears caused by bad shooting angles or weather reasons when manually collecting pictures. When segmenting the wheat ear image, the present invention takes into account that the shape of the wheat ear is regular and approximate to an ellipse. By adaptively calculating the approximate ellipse of the wheat ear, the wheat ear image is further masked with an ellipse on the basis of segmentation. Compared with the traditional method of segmenting the wheat ear image using color features and morphological features, the segmentation effect of the wheat ear can be more effectively improved. When extracting the features of diseased ears, the present invention uses a Resnet network to extract the features of wheat ears. The features output by the Resnet network reflect the local information of the wheat ear image, mainly reflecting the features of the image at certain local positions, which will enhance the distinction between diseased ears and healthy ears of wheat, thereby improving the classification accuracy. Compared with ordinary color features and morphological features, those features only reflect the global features of the image and cannot accurately distinguish between diseased ears and healthy ears of wheat. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flow chart of the present invention

[0011] Figure 2 This is the improved YoloV4 network recognition model structure diagram

[0012] Figure 3 This is a comparison chart between the predicted results and the actual results DETAILED DESCRIPTION

[0013] In order to solve the technical problem, the present invention provides a method for evaluating the occurrence degree of wheat scab in the field based on improved YoloV4. The flow chart of the method is as follows Figure 1 shown.

[0014] A method for evaluating the occurrence degree of wheat fusarium head blight in the field based on improved YoloV4, characterized in that it comprises the following steps:

[0015] S100, photographing images of wheat ears in the observation area;

[0016] S200, using the trained improved YoloV4 target detection network model to detect the group wheat ear images in the observation area and obtain a single wheat ear image;

[0017] S300, counting the number of wheat ears to obtain the total number of wheat ears, and performing feature extraction and random forest classification on the single wheat ear image and counting the number of diseased wheat ears;

[0018] S400, calculating the diseased ear rate of wheat ears in the group according to the total number of wheat ears and the number of diseased wheat ears, and obtaining the degree of occurrence of fusarium head blight in the wheat ears in the group according to relevant national standards;

[0019] Among them, in step S200, the improved YoloV4 target detection network model uses CSPDarknet53 as the feature extraction backbone network, and adds multiple CSP modules to the backbone network. The backbone network CSPDarknet53 connects the SPP structure and the PANet network in sequence. The PANet network outputs the detection box and the confidence within the detection box using the Yolo Head network structure inherited from YoloV3. The Soft-NMS module merges all object confidence boxes and outputs the merged result. The improved YoloV4 network recognition model structure is as follows Figure 2 shown.

[0020] More specifically, in step S200, the improved YoloV4 target detection network model is trained according to the following steps:

[0021] S210, photographing images of wheat ears in a plurality of observation areas;

[0022] S220, under the guidance of a plant protection expert, manually labeling wheat ears in the wheat ear images of the group to obtain a training set;

[0023] S230, input the training set into the improved YoloV4 target detection network model for 100 rounds of training, wherein the learning rate for the first 50 rounds of training is 0.001, and the learning rate after 50 rounds of training is 0.0001;

[0024] S240, saving the improved YoloV4 target detection network model after training.

[0025] More specifically, in step S300, the sum of the numbers of individual wheat ear images is the total number of wheat ears, and the number of diseased wheat ears is obtained according to the following steps:

[0026] S310, preprocessing a single wheat ear image to obtain a rough segmentation result of the wheat ear;

[0027] S320, extracting disease features of wheat ears according to the rough segmentation results of wheat ears;

[0028] S330, using the obtained wheat ear disease features to classify the wheat ear images using a random forest algorithm based on a Cart decision tree and obtain the number of diseased ears.

[0029] More specifically, step S310 includes the following steps:

[0030] S311, segmenting a single wheat ear image using an XYZ color space threshold to obtain a first image, wherein the conversion formula of the XYZ color space is as follows:

[0031]

[0032] Among them, R, G, and B represent the R, G, and B color components of the image respectively;

[0033] S312, using an adaptive segmentation threshold T to segment the single wheat ear image to obtain a second image, where the value of T satisfies the following formula:

[0034] T=max(w0×w1×(μ1-μ0) 2 ),

[0035] Among them, w0 represents the ratio of the number of foreground pixels to the total number of pixels in the image, and its average grayscale is μ0; w1 represents the ratio of the number of background pixels to the total number of pixels in the image, and its average grayscale is μ1;

[0036] S313, performing a logical AND operation on the first image and the second image to obtain a preliminary result of rough segmentation of wheat ears;

[0037] S314, calculate the image center point (x c ,y c ), and then calculate the maximum non-zero pixel and the minimum non-zero pixel (x t ,y t ) and (x b ,y b ), and the maximum non-zero pixel and the minimum non-zero pixel (x l ,y l ) and (x r ,y r ); then calculate |y t -y b | is denoted as a, and |x is calculated l -x r | is denoted as b, and then the single wheat ear image is The region is masked to obtain the final rough segmentation result.

[0038] More specifically, in step S320, the disease characteristics of the wheat ears include multiple or all of the following characteristics:

[0039] S321, using the HSV color space to perform threshold segmentation on the rough segmentation result to obtain the diseased spot area, and then calculating the area ratio of the pixel points to obtain the disease feature A;

[0040] S322, using CMYK color space to perform threshold segmentation on the rough segmentation result to obtain the diseased spot area, and then calculate the area ratio of the pixel points to obtain the disease feature B;

[0041] S323, obtaining disease feature C using the value of the rough segmentation result within the H channel range in the HSV color space;

[0042] S324. Through transfer learning of the ResNet-50 network, the coarse segmentation result is input into the ResNet-50 network to obtain the output result of the coarse segmentation result in the Conv5 layer of the ResNet-50 network. The output result contains the semantic features of a single wheat ear image. The output result is then subjected to principal component analysis (PCA) to extract linearly unrelated variables from the semantic features, followed by maximum pooling to merge linearly unrelated variables, and finally the disease feature D is obtained.

[0043] More specifically, in step S324, the size of the output data of the coarse segmentation result in the Conv5 layer of the ResNet-50 network is 7×7×1024. The PCA principal component analysis is used to obtain an output size of 7×7 and the result is subjected to a maximum pooling operation. The maximum pooling convolution kernel size is 3×3, the step size is 4, and the final size of the disease feature D is 2×2. The Res-Net transfer learning network is trained according to the following steps:

[0044] Prepare multiple single wheat ear pictures, including diseased wheat ear pictures and healthy wheat ear pictures;

[0045] Fill the edges of multiple single wheat ear images with 0 so that their aspect ratio is 1;

[0046] The training set is the image with a pixel size of 224×224 obtained by resampling using bilinear interpolation.

[0047] The Res-Net transfer learning network was trained using the training set for a total of 30 rounds with a learning rate of 0.0001.

[0048] More specifically, in step S320, the disease characteristics include four types: A, B, C, and D. In step S330, the Cart decision tree is established according to the following steps:

[0049] S331. For the data set P of the current node, if the number of samples is less than the threshold or there is no feature, the decision subtree is returned and the recursion of the current node is stopped;

[0050] S332, calculating the Gini coefficient of the sample set P, if the Gini coefficient is less than the threshold, returning to the decision tree subtree, and stopping the recursion at the current node;

[0051] S333, calculating the Gini coefficient of each existing feature of the current node in the data set P;

[0052] S334, among the calculated Gini coefficients of the features in the data set P, select the feature Q with the smallest Gini coefficient and the corresponding eigenvalue q, divide the data set into two parts P1 and P2 according to the optimal feature and the optimal eigenvalue, and simultaneously establish the left and right nodes of the current node, the data set of the left node is P1, and the data set of the right node is P2;

[0053] S335. Recursively call steps S331-S334 on the left and right child nodes to generate a decision tree.

[0054] More specifically, the following steps are also included after step S400:

[0055] S500, the degree of wheat scab incidence in the population is graded according to the following rules based on the diseased ear rate:

[0056] If the diseased ear rate is greater than 0.001 and less than or equal to 0.1, the occurrence degree of wheat fusarium rust is 1;

[0057] If the diseased ear rate is greater than 0.1 and less than or equal to 0.2, the degree of wheat fusarium rust is 2;

[0058] If the diseased ear rate is greater than 0.2 and less than or equal to 0.3, the degree of wheat fusarium rust is 3;

[0059] If the diseased ear rate is greater than 0.3 and less than or equal to 0.4, the degree of wheat fusarium rust is 4;

[0060] If the diseased ear rate is greater than 0.4 and less than or equal to 1.0, the degree of wheat fusarium rust occurrence is 5.

[0061] In order to better explain the positive technical effects obtained by the present invention, wheat ear images collected at the Hefei Academy of Agricultural Sciences in Anhui Province in 2020 were used as verification. 50 images were randomly selected from the collected images as a test set, and the predicted number of wheat ears, predicted diseased ear rate, and predicted occurrence degree using the patented method were recorded respectively and compared with the actual number of wheat ears, actual diseased ear rate, and actual occurrence degree evaluated by plant protection experts. The comparison results are as follows: Figure 3 As shown, PN represents the image number, PWC represents the predicted number of wheat ears, RWC represents the actual number of wheat ears, PDR represents the predicted diseased ear rate, RDR represents the actual wheat ear rate, PDL represents the predicted disease occurrence degree, and RDL represents the actual disease occurrence degree. Figure 3By comparing the difference between the predicted disease occurrence degree and the actual occurrence degree, it can be concluded that the prediction accuracy of the wheat fusarium scaly disease occurrence degree of the method proposed in the present invention is 92%.

[0062] The specific embodiments described in this application are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for evaluating the occurrence degree of wheat fusarium head blight in the field based on improved YoloV4, characterized in that: The steps include: S100, photographing images of wheat ears in the observation area; S200, using the trained improved YoloV4 target detection network model to detect the group wheat ear images in the observation area and obtain a single wheat ear image; S300, counting the number of wheat ears to obtain the total number of wheat ears, and performing feature extraction and random forest classification on the single wheat ear image and counting the number of diseased wheat ears; S400, calculating the diseased ear rate of wheat ears in the group according to the total number of wheat ears and the number of diseased wheat ears, and obtaining the degree of occurrence of fusarium head blight in the wheat ears in the group according to relevant national standards; Among them, in step S200, the YoloV4 target detection network model is improved to use CSPDarknet53 as the feature extraction backbone network, and multiple CSP modules are added to the backbone network. The backbone network CSPDarknet53 is connected to the SPP structure and the PANet network in sequence. The result output by the PANet network uses the Yolo Head network structure inherited from YoloV3 to output the detection frame and the confidence within the detection frame. The Soft-NMS module merges all object confidence frames and outputs the merged result; In step S300, the sum of the numbers of individual wheat ear images is the total number of wheat ears, and the number of diseased wheat ears is obtained according to the following steps: S310, preprocessing a single wheat ear image to obtain a rough segmentation result of the wheat ear; S320, extracting disease features of wheat ears according to the rough segmentation results of wheat ears; S330, using the obtained wheat ear disease features to classify the wheat ear images using a random forest algorithm based on a Cart decision tree and obtain the number of diseased ears; Step S310 includes the following steps: S311, segmenting a single wheat ear image using an XYZ color space threshold to obtain a first image, wherein the conversion formula of the XYZ color space is as follows: Among them, R, G, and B represent the R, G, and B color components of the image respectively; S312, using an adaptive segmentation threshold T to segment the single wheat ear image to obtain a second image, where the value of T satisfies the following formula: T=max(w0×w1×(μ1-μ0) 2 ), Among them, w0 represents the ratio of the number of foreground pixels to the total number of pixels in the image, and its average grayscale is μ0; w1 represents the ratio of the number of background pixels to the total number of pixels in the image, and its average grayscale is μ1; S313, performing a logical AND operation on the first image and the second image to obtain a preliminary result of rough segmentation of wheat ears; S314, calculate the image center point (x c ,y c ), and then calculate the maximum non-zero pixel and the minimum non-zero pixel (x t ,y t ) and (x b ,y b ), and the maximum non-zero pixel and the minimum non-zero pixel (x l ,y l ) and (x r ,y r ); then calculate |y t -y b | is denoted as a, and |x is calculated l -x r | is denoted as b, and then the single wheat ear image is The region is masked to obtain the final rough segmentation result; In step S320, the disease characteristics of wheat ears include multiple or all of the following characteristics: S321, using the HSV color space to perform threshold segmentation on the rough segmentation result to obtain the diseased spot area, and then calculating the area ratio of the pixel points to obtain the disease feature A; S322, using CMYK color space to perform threshold segmentation on the rough segmentation result to obtain the diseased spot area, and then calculate the area ratio of the pixel points to obtain the disease feature B; S323, obtaining disease feature C using the value of the rough segmentation result within the range of the H channel in the HSV color space; S324. Through transfer learning of the ResNet-50 network, the coarse segmentation result is input into the ResNet-50 network to obtain the output result of the coarse segmentation result in the Conv5 layer of the ResNet-50 network. The output result contains the semantic features of a single wheat ear image. The output result is then subjected to principal component analysis (PCA) to extract linearly unrelated variables from the semantic features, followed by maximum pooling to merge linearly unrelated variables, and finally the disease feature D is obtained.

2. The method for evaluating the occurrence degree of wheat fusarium head blight in the field based on improved YoloV4 according to claim 1, characterized in that: In step S200, the improved YoloV4 target detection network model is trained according to the following steps: S210, photographing images of wheat ears in a plurality of observation areas; S220, under the guidance of a plant protection expert, manually labeling wheat ears in the wheat ear images of the group to obtain a training set; S230, input the training set into the improved YoloV4 target detection network model for 100 rounds of training, wherein the learning rate for the first 50 rounds of training is 0.001, and the learning rate after 50 rounds of training is 0.0001; S240, saving the improved YoloV4 target detection network model after training.

3. The method for evaluating the occurrence degree of wheat fusarium head blight in the field based on improved YoloV4 according to claim 1, characterized in that: In step S324, the size of the output data of the coarse segmentation result in the Conv5 layer of the ResNet-50 network is 7×7×1024. The PCA principal component analysis is used to obtain an output size of 7×7 and the result is subjected to a maximum pooling operation. The convolution kernel size of the maximum pooling is 3×3, the step size is 4, and the final size of the disease feature D is 2×2. The Res-Net transfer learning network is trained according to the following steps: Prepare multiple single wheat ear pictures, including diseased wheat ear pictures and healthy wheat ear pictures; Fill the edges of multiple single wheat ear images with 0 so that their aspect ratio is 1; The training set is the image with a pixel size of 224×224 obtained by resampling using bilinear interpolation. The Res-Net transfer learning network was trained using the training set for a total of 30 rounds with a learning rate of 0.0001.

4. The method for evaluating the occurrence degree of wheat fusarium head blight in the field based on improved YoloV4 according to claim 1, characterized in that: In step S320, the disease characteristics include four types: A, B, C, and D. In step S330, the Cart decision tree is established according to the following steps: S331. For the data set P of the current node, if the number of samples is less than the threshold or there is no feature, the decision subtree is returned and the recursion of the current node is stopped; S332, calculating the Gini coefficient of the sample set P, if the Gini coefficient is less than the threshold, returning to the decision tree subtree, and stopping the recursion at the current node; S333, calculating the Gini coefficient of each existing feature of the current node in the data set P; S334, among the calculated Gini coefficients of the features in the data set P, select the feature Q with the smallest Gini coefficient and the corresponding eigenvalue q, divide the data set into two parts P1 and P2 according to the optimal feature and the optimal eigenvalue, and simultaneously establish the left and right nodes of the current node, the data set of the left node is P1, and the data set of the right node is P2; S335. Recursively call steps S331-S334 on the left and right child nodes to generate a decision tree.

5. The method for evaluating the occurrence degree of wheat fusarium head blight in the field based on improved YoloV4 according to claim 1, characterized in that: After step S400, the following steps are also included: S500, the degree of wheat scab incidence in the population is graded according to the following rules based on the diseased ear rate: If the diseased ear rate is greater than 0.001 and less than or equal to 0.1, the degree of wheat fusarium rust is 1; If the diseased ear rate is greater than 0.1 and less than or equal to 0.2, the degree of wheat fusarium rust is 2; If the diseased ear rate is greater than 0.2 and less than or equal to 0.3, the degree of wheat fusarium rust is 3; If the diseased ear rate is greater than 0.3 and less than or equal to 0.4, the degree of wheat fusarium rust is 4; If the diseased ear rate is greater than 0.4 and less than or equal to 1.0, the degree of wheat fusarium rust occurrence is 5.