CN114199880A

CN114199880A - Citrus disease and insect pest real-time detection method based on edge calculation

Info

Publication number: CN114199880A
Application number: CN202111454283.3A
Authority: CN
Inventors: 姚丽萍; 戴卓辰; 彭爽; 谢守勇; 张军辉
Original assignee: Southwest University
Current assignee: Southwest University
Priority date: 2021-11-22
Filing date: 2021-11-22
Publication date: 2022-03-18

Abstract

The invention discloses a citrus disease and insect pest real-time detection method based on edge calculation. The real-time detection method comprises the steps of collecting hyperspectral atlas data of citrus leaves by a hyperspectral imaging system, analyzing to obtain main characteristic wave bands of plant diseases and insect pests, using the characteristic wave bands as the center wavelength of an optical filter selected by a multispectral camera, extracting a single-channel image of the citrus leaves under the characteristic wave bands, constructing an image data set, training and evaluating a deep learning model, deploying model weights into an FPGA board card through quantization and compiling processing, reading a spectral image of the citrus leaves collected by the multispectral camera on site, and outputting inference results to an FPGA board card display screen to detect whether the plant diseases and insect pests exist in citrus seedlings. The invention combines the multispectral detection technology, the deep learning method and the edge calculation for the first time and is used for the on-site real-time detection of the citrus diseases and insect pests. Not only improve the speed of detecting oranges and tangerines plant diseases and insect pests, but also guaranteed the precision of detection.

Description

Citrus disease and insect pest real-time detection method based on edge calculation

Technical Field

The invention relates to the field of agricultural pest detection, in particular to a citrus pest real-time detection method based on edge calculation.

Background

In the production of oranges, symptoms such as magnesium deficiency, iron deficiency and the like often occur. Research reports indicate that once the citrus lacks magnesium, the leaves of the citrus are green, the fruits of the citrus are generally small, and the quality and the yield of the citrus are reduced. When the citrus is lack of iron, the leaves are yellow; with the deepening of symptoms, moderate iron deficiency can also cause the reduction of the yield of oranges, the reduction of the sweetness of fruits, the increase of acidity, small fruits, light color of peels and other symptoms; the phenomenon of flower and fruit drop is increased due to serious iron deficiency, and the yield is reduced finally; when the citrus yellow shoot disease occurs, the leaves of the citrus yellow shoot usually show symptoms of yellowing, withering and the like on the outside, and the leaves usually show symptoms of pigment damage, increased antioxidant enzyme activity, changed nutrient absorption, abnormal sugar metabolism and the like on the inside, so that the scattering and absorption of light by the citrus leaves are influenced. The spectrum detection method generally uses a hyperspectral imaging system to image visible light-near infrared wave bands (400-1000 nm) of leaves of a detected plant, and in the wave band range of imaging, presented spectral characteristics can better reflect information such as pigments, water, dry matters and the like in the plant, so that citrus diseases and insect pests are analyzed through citrus leaves.

Aiming at the identification of citrus diseases and insect pests, the traditional field diagnosis is mainly carried out in a manual visual inspection mode, and detection personnel comprehensively judge according to different characteristics expressed by citrus leaves. The method may cause a low detection accuracy due to factors such as misjudgment and non-uniform standard. The spectrum detection method mainly adopts a hyperspectral imaging system to image citrus leaves, and the obtained spectrum image can better reflect the internal information of plants so as to analyze whether citrus seedlings are sick or not.

With the rapid development of computer technology, in recent years, relevant scholars combine smart phones with machine learning methods, and some specific citrus diseases and insect pests can be detected on site in real time at present, and certain results are obtained. However, the research reports adopt a common RGB visible light camera to carry out imaging, so as to identify the characteristics of the citrus leaves, such as form, color, lines and the like, and analyze whether the plants are diseased. However, since the wavelength band received by the common RGB visible light camera is close to the spectrum wavelength band of the image perceived by human eyes, when the citrus leaves show similar symptoms of deficiency of elements and yellow dragon infection, the similar misjudgment of human eyes can be caused. Aiming at the current situation, how to accurately and quickly find whether the citrus is diseased or not in the growth process of the citrus and timely remedy the diseased citrus is a key problem in improving the yield and quality of the citrus, and has very important significance for the development of citrus industry in China and even the whole world. Based on the method, the multi-spectral imaging technology and the method combining deep learning and edge calculation are adopted to research the citrus disease and insect pest characteristics, and the model obtained by deep learning training is deployed on the FPGA board card to carry out real-time model reasoning, so that the detection equipment capable of analyzing the citrus disease and insect pest on site in real time is realized.

Disclosure of Invention

The invention aims to provide a method for detecting and analyzing citrus diseases and insect pests in real time on site.

The invention provides a field analysis method considering both detection precision and detection speed, aiming at the problems of low precision, non-uniform standard and the like in the traditional field detection method of citrus diseases and insect pests.

In order to achieve the purpose, the invention adopts the following technical scheme:

a citrus disease and insect pest real-time detection method based on edge calculation comprises the following steps:

step 1: collecting diseased citrus leaves and healthy citrus leaves of different varieties, different places and different growth periods, cleaning and sorting the diseased citrus leaves and the healthy citrus leaves, and then scanning and imaging the diseased citrus leaves and the healthy citrus leaves by using a hyperspectral imaging system to obtain hyperspectral map data of a sample;

step 2: in order to reduce the influence of uneven light source intensity under different wave bands, irregular samples and dark current in an imaging lens, black and white correction is carried out on the hyperspectral atlas data;

and step 3: importing the corrected hyperspectral atlas data into analysis software, dividing an interesting area, and extracting spectral curve data of all samples;

and 4, step 4: preprocessing the spectral curve data by adopting a multivariate scattering correction algorithm to obtain spectral curve correction data;

and 5: analyzing the spectral curve correction data by adopting a recursive characteristic elimination algorithm, and screening out a plurality of main characteristic wave bands with the maximum classification weight;

step 6: extracting the single-channel gray image under the characteristic wave band by adopting an edge filling and proportional scaling method to obtain an image data set meeting the requirement of a depth model;

and 7: dividing the image data set into a training set, a verification set and a test set according to the ratio of 6: 2;

and 8: building a deep learning model at a computer end, introducing the image data set, performing model training and reasoning after data enhancement, and storing a trained model weight file;

and step 9: loading the model weight file to perform model reasoning at a computer end, and analyzing whether the reasoning precision meets the design requirement;

step 10: carrying out quantitative compiling on the model weight files meeting the requirements to obtain special model weight files which can be used by the FPGA board card;

step 11: loading the special model weight file into an FPGA board card memory;

step 12: the multispectral camera is connected with the FPGA board card by using a data line, the multispectral camera is driven by the FPGA board card to shoot a citrus blade sample, and a multispectral image of the citrus blade is obtained and read into an internal memory of the FPGA board card;

step 13: processing the multispectral image in the FPGA memory through an edge filling and scaling algorithm to obtain a processed image;

step 14: the FPGA board card calls the special model weight file from the memory, and model reasoning is carried out on the preprocessed image to obtain a model reasoning result;

step 15: outputting the inference result to a display screen of the FPGA board card for displaying;

step 16: repeating the steps 12-15 to obtain a plurality of model reasoning results;

and step 17: carrying out model accuracy verification by analyzing the multiple model reasoning results, and judging whether the accuracy of the special model weight file meets the design requirements or not;

step 18: in the field real-time monitoring, the steps 12-15 are repeated for the citrus seedlings to be detected, so that the purpose of detecting the citrus diseases and insect pests in real time on the field is achieved.

In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:

(1) the invention combines the deep learning method, the edge calculation technology and the spectral analysis technology for the first time, and is applied to the real-time detection of the citrus diseases and insect pests.

(2) According to the method, the FPGA board card is used as an edge computing platform, a deep learning model is deployed, so that the spectral image is analyzed, the platform has a higher energy efficiency ratio under the same power consumption, and can be used as a low-power-consumption handheld device on site, and the applicability and the popularization of the method are greatly improved.

(3) The method for characteristic band extraction, data set establishment, deep learning training, model deployment and the like can be applied to other types of spectral analysis problems, and is not limited to real-time detection of citrus diseases and insect pests.

Drawings

FIG. 1 is a schematic flow chart of the method of the present invention

FIG. 2 is a schematic diagram of the method for detecting citrus greening disease in real time

FIG. 3 is a comparison graph of the test result precision obtained by the method at the computer end and the FPGA board card

FIG. 4 is a comparison graph of single inference time obtained by the method at the computer end and the FPGA board card

FIG. 5 is a sub-flowchart of the FPGA board card performing real-time edge calculation according to the present invention

Detailed Description

In order to make the technical problems, technical solutions and advantageous effects to be solved by the present invention more apparent, the present invention is described in detail below with reference to the embodiments. It should be noted that the specific embodiments described herein are only for illustrating the present invention and are not to be construed as limiting the present invention, and products that can achieve the same functions are included in the scope of the present invention.

The present invention will be described in detail below.

Referring to the attached figure 1, the citrus disease and insect pest real-time detection method based on edge calculation comprises the following steps;

step 1: collecting citrus leaves of diseased citrus leaves and healthy citrus leaves of different varieties, different places and different growth periods, cleaning and sorting the citrus leaves, and then scanning and imaging by using a hyperspectral imaging system to obtain hyperspectral map data of the sample;

step 11: loading the special model weight file into an FPGA board card memory;

Example one

The embodiment is used for detecting the huanglongbing disease in the citrus diseases and pests, and the detection is mainly carried out by using the characteristic difference of the spectrum and the image information of different properties of different wave bands of the leaves of the citrus seedlings after the citrus seedlings are infected with the huanglongbing disease.

The citrus greening disease real-time detection method based on edge calculation is shown in fig. 2, and the specific process is as follows:

firstly, performing hyperspectral imaging on healthy citrus leaves and citrus leaves infected with yellow dragon disease germs by a hyperspectral imaging system to obtain hyperspectral map data, then performing black-and-white correction, importing the corrected hyperspectral map data into analysis software, dividing regions of interest, extracting spectral curve data of all samples, and then preprocessing the spectral curve data by adopting a multivariate scattering correction algorithm to obtain spectral curve correction data; then, analyzing the spectral curve correction data by adopting a recursive characteristic elimination algorithm to obtain a plurality of main characteristic wave bands with the maximum classification weight; extracting the single-channel gray image under the characteristic wave band by adopting an edge filling and proportional scaling method to obtain an image data set meeting the requirement of a depth model;

dividing the image data set into a training set, a verification set and a test set according to the ratio of 6: 2;

and then building a deep learning classification model, performing data enhancement on the image data set, then performing model training and reasoning, and storing the trained model weight file. Loading the model weight file and carrying out reasoning at a computer end, and analyzing whether the reasoning precision meets the design requirement; carrying out quantitative compiling on the model weight files meeting the requirements to obtain special model weight files which can be used by the FPGA board card, and loading the special model weight files into an internal memory of the FPGA board card;

the multispectral camera is connected with the FPGA board card by using a data line, the multispectral camera is driven by the FPGA board card to shoot a citrus blade sample, a multispectral image of the citrus blade is obtained, and the multispectral image is read into an internal memory of the FPGA board card for preprocessing; calling the special model weight file by the FPGA board card, and reasoning the preprocessed multispectral image to obtain a model reasoning result; the FPGA board card outputs the inference result to a display screen of the FPGA board card for displaying;

when the citrus greening disease is diagnosed on site, the method is used for shooting and imaging the leaves of the citrus seedlings to be detected, carrying out model reasoning, outputting results and the like. So as to analyze and determine whether the citrus yellow dragon disease exists in the shooting sample.

In order to verify the model inference precision and speed of the FPGA end, the computer end loads data shot by the multispectral camera, and adopts a model weight file before quantization and compilation to perform performance verification comparison, wherein the overall inference precision is shown in figure 3, and the single model inference time is shown in figure 4. In the reasoning process, the overall power of the computer end is more than 300w, and the power of the FPGA board card is only 18 w. Therefore, in the embodiment, the FPGA board card is used for edge calculation reasoning, the energy efficiency ratio is better than that of a computer end, and the precision and the speed meet the design requirements.

The above description is only for the preferred embodiment of the present invention, and should not be used to limit the scope of the claims of the present invention. While the foregoing description will be understood and appreciated by those skilled in the relevant art, other equivalents may be made thereto without departing from the scope of the claims.

Claims

1. A citrus disease and insect pest real-time detection method based on edge calculation is characterized by comprising the following steps:

step 1: respectively scanning and imaging diseased citrus leaves and healthy citrus leaves of different varieties, different places and different growth periods by using a hyperspectral imaging system to obtain hyperspectral map data of the sample;

step 2: performing black-and-white correction on the hyperspectral atlas data;

and step 3: importing the corrected hyperspectral map data into analysis software, dividing regions of interest, and extracting spectral curve data of all samples;

step 6: extracting a single-channel gray image under the characteristic wave band by adopting an edge filling and scaling method to obtain an image data set meeting the requirement of a depth model, and dividing the image data set into a training set, a verification set and a test set according to the ratio of 6: 2;

and 7: and (3) building a deep learning model at a computer end, introducing the image data set, performing model training and reasoning after data enhancement, and storing a trained model weight file. Loading the model weight file and carrying out reasoning at a computer end, and analyzing whether the reasoning precision meets the design requirement;

and 8: carrying out quantitative compiling on the model weight files meeting the requirements to obtain special model weight files which can be used by the FPGA board card;

and step 9: loading the special model weight file into an FPGA board memory to perform real-time model reasoning;

step 10: the multispectral camera is connected with the FPGA board card by using a data line, the multispectral camera is driven by the FPGA board card to shoot a citrus blade sample, a multispectral image of the citrus blade is obtained, and the multispectral image is read into an internal memory of the FPGA board card for preprocessing;

step 11: calling the special model weight file by the FPGA board card, and reasoning the preprocessed multispectral image to obtain a model reasoning result;

step 12: the FPGA board card outputs the inference result to a display screen of the FPGA board card for displaying;

step 13: repeating the steps 10 to 12, carrying out accuracy verification on the image data set, and judging whether the method meets the designed precision requirement;

step 14: in the field real-time monitoring, the steps 10 to 12 are repeated for the citrus seedlings to be detected, so that the purpose of detecting the citrus diseases and insect pests in real time on the field is achieved.

2. The method according to claim 1, wherein the recursive feature elimination algorithm used in step 5 is used to analyze and screen spectral information of citrus diseases and insect pests to find the dominant characteristic bands of diseased citrus fruit.

3. The method as claimed in claim 1, wherein the deep learning model for citrus pest detection built in the step 7 is an image classification network model based on a convolutional neural network as a main body, and input data of the deep learning model is a preprocessed citrus blade single-channel gray scale map.

4. The method as claimed in claim 1, wherein the data enhancement work taken on the citrus blade data set in step 7 includes processes of flipping, rotating, shifting, adding gaussian noise to the image data at random, and the like, aiming to improve generalization capability of the trained model in reasoning.

5. The method according to claim 1, wherein the step 10 of preprocessing the read-in images of citrus fruit leaves includes processing the read-in images using scaling and edge filling to ensure that data dimensions meet model input requirements without image distortion.

6. The method according to claim 1, wherein the reasoning result finally output to the display screen in the step 12 is the current detected citrus leaf image, and the type and probability of the disease.