Research On Online Detection Method Of Combine Operation Quality Based On Vision

Combine harvester is the main agricultural machinery equipment of rice harvest mechanization.In recent years,agricultural machinery has been developed rapidly in the intelligent,which greatly improves the efficiency and quality of agricultural production.In this paper,the necessity and significance of on-line detection of combine operation quality have been introduced according to the current situation of intelligent development of combine at home and abroad.The research progress of on-line detection of combine operation quality at home and abroad has been summarized,and a vision based on-line detection system of combine operation quality has been proposed.Using image machine vision method,the system could detect the rice impurity rate and broken rate data online,and intuitively feed them back to the driver through the display terminal,so as to provide data reference for the reasonable adjustment of combine operation parameters and ensure that the combine operation state could keep in a better working state;at the same time,the operation data is stored in network to facilitate remote monitoring At the same time,it is the basis for further research and improvement of the algorithm.The main contents of this paper are as follows:Firstly,the online detection method of combine operation quality was analyzed.By analyzing the influence factors of combine operation process on operation quality,the research object of this paper has been determined as rice impurity rate and broken rate.The characteristics of intact rice,rice impurities and broken rice in combine grain box were analyzed to determine the main characteristic basis of the detection method described in this paper.Next,detection method of rice impurity rate based on the improved Mask RCNN(Mask regions with Convolutional Neural Networks).The residual module in RESNET has been improved by topology to improve the network recall rate;the anchor size of RPN layer has been optimized according to the characteristics of stem impurities,and the anchor selection method has been optimized;the ROI align layer has been optimized by double integral method according to the image adhesion;finally,the edge loss calculation has been introduced to solve the problem of poor edge segmentation accuracy.Image augmentation technology has been used to expand the image samples to solve the problem of lack of training samples.The comprehensive evaluation index F1 has been selected to measure the recognition accuracy and recall rate.The experiment showed that the recognition rate of this method could reach 91.12%,which indicates that this method can effectively identify stem impurities in rice images.Thirdly,the detection method of rice broken rate based on convolutional neural network was proposed Firstly,the contrast of the collected rice image has been enhanced,and then a simple linear iterative clustering algorithm has been used to pre-segment it.Based on the pre-segmentation results,convolutional neural network has been used to train it to obtain the segmented image.Finally,the broken rice has been extracted based on the HSV color space.Using this method to identify and segment 100 rice images,the results showed that the recognition rate F1 of broken rice could reach 89.11%,which indicated that this method can effectively recognize broken rice in rice images.Finally,verify the performance of the impurity rate and broken rate detection system through experiments.The detection method of impurity rate and broken rate described in this paper has been systematically integrated,including image acquisition module,image processing module,pixel quality ratio conversion module,data display and storage module.The accuracy and stability of the method were verified by experiments.The experimental results showed that this method has a certain recognition accuracy in the rice image with 0.5% ～ 1% impurity rate and broken rate,and has a high recognition accuracy in the rice image with 1.5% ～ 2.5% impurity rate and broken rate.Finally,in the bench test of simulating the actual detection environment,the average absolute percentage error between the detection data of impurity rate and broken rate with the actual data was 13.33% and 14.36% respectively,which meets the expected detection requirements of the system.