| Soybean is an important oil crop in my country,and mechanized harvesting is an important part of green and efficient soybean production.Soybean has the agronomic characteristics of low pod formation and easy breakage.Improper setting of machine parameters can easily lead to high breakage rate and impurity content.At present,there is a lack of real-time detection methods for the brokenness and impurities in the mechanized soybean harvesting process.Drivers need to stop when they understand the harvesting quality.Relying on visual judgment,they cannot find problems in the harvesting process in time,resulting in uneven harvesting quality and affecting economic benefits.In this paper,the real-time monitoring method of soybean crushing rate and impurity content in the mechanized harvesting process is studied.Machine vision and deep learning technology are used to monitor soybean crushing rate and impurity content in real time.A soybean component image segmentation and recognition algorithm based on improved Deep Lab V3+ and U-Net deep learning network combined with attention was proposed,and a quantitative model of impurity rate and broken rate based on image information was established,and the broken rate and impurity rate of soybean harvested mechanized were realized.The online monitoring provides data support for the intelligent control of the parameters of the driver and the combine harvester.The main research contents and conclusions of this paper are as follows:1)Research on the online detection method of machine-harvested soybean crushing rate and impurity content based on machine vision.The main reasons for the broken grains and impurities during the operation of the soybean combine harvester were analyzed.According to the existing manual detection methods of broken rate and impurity rate,a method for detecting broken rate and impurity rate of mechanized harvested soybean based on image information is proposed.Using opencv to count impurities and the pixel area of soybean grains in the sample image,the coupling relationship between soybean grain quality and pixel area was established,and a linear regression calculation model was fitted.The regression analysis results showed the correlation coefficient;A univariate linear regression calculation model is fitted,and the regression analysis results show the correlation coefficient.Combined with the manual detection of the broken rate and the calculation method of the impurity rate,the calculation method of the broken rate and the impurity rate based on the image information is formulated.2)Developed an on-line detection device for the broken rate and impurity content of soybean mechanized harvesting.The device consists of an industrial computer,a grain collection device,an industrial computer,an industrial camera,an LED light source,and a steering gear.The grain collection device is installed under the grain outlet of the soybean combine harvester to collect soybeans falling into the grain sampling tank.The industrial computer collects soybean images through an industrial camera and performs image processing,identification and display.The lower computer controls the motor action.Realize the dynamic sampling of soybean sample images in the mechanized harvesting process.3)A soybean component image segmentation and recognition algorithm based on improved Deep Lab V3+ and U-Net deep learning network combined with attention is proposed.Labelme software was used to manually label the image components of the collected soybean samples(intact grains,broken grains,impurities)to establish a data set required for deep learning algorithm training;an improved Deep Lab V3+ and an attention-combined U-Net network structure were designed.Build a deep learning training environment,complete the training,testing and verification of the model offline,and obtain the optimal deep learning model.The test results show that: based on the improved Deep Lab V3+ model,the precision rate of complete grains is 95.34%,the recall rate is 90.45%,the comprehensive evaluation index F1 is 92.81%,the precision rate of broken grains is 83.89%,the recall rate is 86.48%,and the comprehensive evaluation index F1 is 84.67 %,the impurity precision rate is 97.82%,the recall rate is 77.32%,and the comprehensive evaluation index F1 is 86.19%;the UNet model based on combined attention has a complete grain precision rate of 93.34%,a recall rate of 96.96%,and a comprehensive evaluation index F1.It is 94.87%,the precision rate of broken grain is 97.23%,the recall rate is 88.35%,the comprehensive evaluation index F1 is 92.50%,the impurity precision rate is 94.65%,the recall rate is90.61%,and the comprehensive evaluation index F1 is 92.37%;Compared with the Depp Lab V3+ network model,the U-Net network based on combined attention predicts that the image segmentation effect is significantly better,with fewer missed and incorrectly segmented areas,and the comprehensive evaluation index F1 of complete grains is increased by 2.06% on average,and the comprehensive evaluation index of broken grains.7.83% and the impurity comprehensive evaluation index F1 increased by 6.18% on average.Therefore,this paper adopts the U-Net network based on combined attention as the detection algorithm of the image processing part of the online detection device for broken rate and impurity rate.4)Experiments to verify the performance of the on-line detection device for the broken rate and impurity content of soybean mechanized harvesting.The indoor bench test results show that the average time of single detection of the device is 12 s,the maximum value of impurity rate detection is 1.91%,the minimum value is 0.15%,and the average value is 0.78%;the maximum value of breakage rate is 5.33%,and the minimum value is 1.51%.%,with an average value of 3.17%;compared with the manual detection,the average error of the soybean crushing rate and impurity content detection device was 0.13%,and the average error of the impurity content rate was0.25%.The field test results show that the soybean crushing rate and impurity content detection device works normally,and realizes the dynamic online detection of soybean samples;%;the maximum value of breakage rate is 7.42%,the minimum value is1.51%,and the average value is 3.60%;the average value of error with manual detection is 0.17%,and the average value of error of impurity rate is 0.1%.Therefore,the method developed in this paper can successfully complete the online detection of the harvest quality when the combine harvester is operating. |
