CN-115661491-B - Monitoring method for pest control in tea tree planting

CN115661491BCN 115661491 BCN115661491 BCN 115661491BCN-115661491-B

Abstract

The invention relates to a monitoring method for pest control in tea tree planting, which comprises the following steps of a, firstly, carrying out feature extraction on an image sample by adopting a main network, simultaneously selecting shallow layer features of the image sample to carry out regression operation, b, gradually adding deep layer features in each layer of the main network to the shallow layer features in an up-sampling and convolution layer combination mode, c, merging each layer of deep layer features of the main network with corresponding shallow layer features after each layer is overlapped to serve as an output layer of a neural network, finishing regression identification on tea leafhoppers, and simultaneously, finishing repeated identification de-duplication treatment on the tea leafhoppers at the same position on different output layers according to identification and regression operation results and combining the defined yellow plate range, and finally carrying out quantity calculation. The monitoring method can efficiently identify and count the tea leafhoppers, has high judgment accuracy, saves time and labor and saves labor cost.

Inventors

CHEN SHICHUN
WANG XIAOQING
JIANG HONGYAN
HU XIANG
PENG PING

Assignees

重庆市农业科学院

Dates

Publication Date: 20260512
Application Date: 20200915

Claims (4)

1. The monitoring method for pest control in tea tree planting is characterized by comprising the following steps of: a. firstly, adopting a standard VGG16 network structure as a main network, extracting features of an image sample containing all outlines of a yellow panel, and simultaneously selecting shallow features of the image sample for regression operation; b. The deep features containing abundant semantic information in each layer of the main network are stacked on the shallow features layer by layer in a combination mode of up-sampling and convolution layers for enriching the semantic information of the shallow features, wherein in the step b, the up-sampling adopts a mode of directly assigning corresponding positions and filling positions to be zero, namely if data with the sampling width of N and the sampling height of M are up-sampled A, B times respectively, the data with the width of NA and the sampling height of MB are obtained, wherein (NA+ ,MB+ ) The corresponding positions of points in the original image are removed, and the rest points after the corresponding positions of points in all the original image are removed are filling positions; c. The method comprises the steps of merging deep features of each layer of a main network with corresponding shallow features of each layer after superposition, carrying out regression identification on tea leafhoppers on the output layer of the neural network as an output layer of the neural network, carrying out identification on the tea leafhoppers at different positions and de-duplication treatment on the tea leafhoppers at the same position on the same output layer in the regression identification process, and simultaneously, carrying out de-duplication treatment on repeated identification on the tea leafhoppers at the same position on different output layers according to the identification and regression operation results of the tea leafhoppers and combining the delimited yellow plate range, and finally carrying out quantity calculation on the tea leafhoppers on the yellow plate; the training method of the network adopts a general recognition model iterative training mode, a training set is obtained by adopting a manual labeling mode of selecting targets on different data, the size is estimated by adopting a template matching mode in the step a, and therefore the optimal characteristic layer in a main network is selected as an input layer of a shallow network, and the template matching mode specifically comprises the following steps: Firstly, drawing square frames with the side length of h and the unit of cm at four corners and the center of a yellow board, carrying out template matching operation by adopting rectangular frames in original pictures of images, respectively determining four vertex angles and centers of the yellow board, and then simultaneously calculating the number of pixel points contained in the rectangular frames in the matching of the four vertex angles and the centers, namely n 1 、n 2 、n 3 、n 4 and n 0 , respectively estimating the number of imaging pixel points of the tea leafhoppers in the original pictures of the images, wherein the specific formula is as follows: = wherein i=0, 1,2,3,4; wherein k is the length of a tea leafhopper body and the unit is millimeter, 1 is millimeter, h is the side length of a drawn square frame and the unit is centimeter; the rectangular yellow plate is divided into four triangular areas through four vertex angles and a center, and then the pixel size occupied by the average tea leafhoppers in each area is calculated, wherein the specific formula is as follows: Wherein I=1, 2,3,4, and ; And finally, selecting an optimal characteristic layer for identifying the tea leafhoppers according to the pixel size occupied by the tea leafhoppers in each area, wherein the specific formula is as follows: = wherein i=1, 2,3,4; In the formula, And k represents selecting a characteristic layer output by a kth block as an input layer of a shallow neural network, namely a tea lesser leafhopper identification layer of the area.
2. The method for monitoring pest control in tea tree planting according to claim 1, wherein the repeated identification of whether the same position is used in the step c is performed by comparing the ratio of the intersection area and the union area of two identification frames at the same position with a threshold value.
3. The method for monitoring pest control in tea tree planting according to claim 1 or 2, wherein the defining of the yellow board range in the step c is determined according to the matching result of rectangular frames on four top corners of the yellow board or the image segmentation mode.
4. The method for monitoring pest control in tea tree planting according to claim 3, wherein in the step a, an image sample containing all outlines of yellow boards is an image frame obtained by any one of high-definition camera shooting or mobile phone shooting, and is converted into a format directly read by a deep learning frame through preprocessing and training, and in the step a, regression operation is performed by adopting a set anchor ratio of 1:k or k to 1, wherein k is the body length of tea leafhoppers.

Description

Monitoring method for pest control in tea tree planting The application discloses a division application of 'a method for identifying and counting tea leafhoppers based on a convolutional neural network', aiming at the application number 202010967815.2. Technical Field The invention relates to the technical field of pest control in tea gardens, in particular to a method for monitoring pest control in tea tree planting. Background The tea leafhoppers are one of important pests which are most widely distributed in various tea areas in China and seriously damaged and influence the yield and quality of tea. In the middle and downstream tea areas of the Yangtze river of China, the loss caused by the tea is about 10% -15% of the summer and autumn tea in the normal year, and the serious disaster year is as high as more than 50%. In the pest control of tea gardens, one of the primary tasks is the control of the tea lesser leafhoppers, but the control measures are taken for the tea lesser leafhoppers, and the occurrence number and the occurrence trend of the tea lesser leafhoppers in the tea gardens are required to be monitored. At present, a manual monitoring method is mainly adopted, namely investigation is carried out on the whole day of the sunny morning dew without drying or on the whole day of the cloudy day, the number of insect mouths on 100 tender leaves (30 bud tips) is randomly investigated, but the method has higher requirements on professional knowledge, judgment experience (the tea leafhopper is agile and active in activity and climbs) and vision level (the tea leafhopper is smaller), and the screening, the identification and the counting are carried out manually for a long time, so that time and labor are wasted. Meanwhile, the accuracy is greatly influenced by artificial factors, the fluctuation range and the error of each identification are large, and the monitoring of the tea leafhoppers cannot be accurately, quickly and effectively finished. Disclosure of Invention The invention aims to provide a monitoring method for pest control in tea tree planting, which can efficiently finish the identification and counting of tea leafhoppers, and has the advantages of high judgment accuracy, time saving, labor saving and labor cost saving. The aim of the invention is achieved by the following technical scheme: the monitoring method for pest control in tea tree planting is characterized by comprising the following steps of: a. firstly, adopting a standard VGG16 network structure as a main network, extracting features of an image sample containing all outlines of a yellow panel, and simultaneously selecting shallow features of the image sample for regression operation; b. The deep features containing abundant semantic information in each layer of the main network are stacked on shallow features layer by layer in a combination mode of up-sampling and convolution layers for enriching the semantic information of the shallow features, wherein in the step b, the up-sampling adopts a mode of directly assigning corresponding positions and filling positions to be zero, namely if data with the sampling width of N and the sampling height of M are up-sampled A, B times respectively, the data with the width of NA and the sampling height of MB are obtained, wherein the data with the sampling width of NA and the sampling height of MB are obtained by the up-sampling method The corresponding positions of points in the original image are removed, and the rest points after the corresponding positions of points in all the original image are removed are filling positions; c. And simultaneously, according to the identification and regression operation results (the regression operation result comprises the position information of the identified tea leafhoppers) of the tea leafhoppers, combining the defined yellow plate range, completing the repeated identification and duplication removal treatment of the tea leafhoppers at the same position on different output layers, and finally carrying out the quantity calculation of the tea leafhoppers on the yellow plates. The training method of the network adopts a general recognition model iteration training mode, a training set is obtained by adopting a manual labeling mode of selecting targets on different data, in order to weaken the influence of the distance randomness of imaging pixels and shooting distances of cameras, the size is estimated by adopting a template matching mode in the step a, thereby selecting the best characteristic layer in a main network as an input layer of a shallow network, and the template matching mode comprises the following specific steps: Firstly, drawing square frames with the side length of h and the unit of cm at four corners and the center of a yellow board, carrying out template matching operation by adopting rectangular frames in original pictures of images, respectively determining four vertex angles and centers of the yellow board, and then simultaneo