| Hybrid rice cultivation emphasizes the requirements of less input and strong seedlings,and uses its tillering ability to increase yield,but the existing rice seeding technology was difficult to meet its agronomic requirements,resulting in the lack of highyield advantages and restricting the development of rice plantation in China.In this paper,an intelligent constant-quantity seeding device for hybrid rice based on embedded system and machine vision was developed,which effectively improves the seeding performance of hybrid rice and lays the foundation for future hybrid rice mechanization.The device was installed on the seedling production line,and can realize the online detection of seeding performance of hybrid rice seedlings.The test results were fed back to the control system of the device.The control system regulates the device timely according to the detection results to achieve intelligent constant seeding.At the same time,a method for detecting seeding performance of hybrid rice pot-tray seedlings based on deep learning was also proposed,which realized more accurate detection of seedling performance of pot-tray seedlings.The main contents and findings are as follows:(1)An intelligent constant-quantity seeding device for hybrid rice based on embedded system and machine vision was designed based on the embedded system,which make the idea of intelligent constant-quantity seeding technology come true.Based on the vibratory precision seeder,an embedded system was designed.A machine learning model was applied in the embedded system.Machine vision and control technology were used to realize the online detection and seeding-quantity control of the seedling performance of blanket-type tray and pot tray.Real-time collection of seed images after seeding on the rice seedling production line,then processing and analysis the seed images to obtain the seeding-quantity information of seeding tray,and comparison with the set value of seeding quantity to obtain the seeding-quantity deviation.The device adjusted the rotation speed of the seeder’s seeding wheel according to the deviation,to realize the intelligent closed-loop control of the seeding quantity,and ensure the intelligent constant-quantity seeding.(2)Three machine learning models of random forest,BP neural network and support vector machine were used to study the seeding performance detection and control methods of hybrid rice blanket-type seedlings.The BP neural network was selected as the prediction model of seeding quantity to achieve the accuracy of detection and the effectiveness of regulation in the condition of low seeding quantity.The methods of image acquisition and image format conversion were elaborated in detail,and an image processing algorithm was used to detect the presence of seedling trays and determine the effective image.By preprocessing the seeding tray image,the seed image was segmented from the seedling tray image.The shape features of the connected regions of the seeds was extracted to train the BP neural network.The trained BP neural network was imported into the embedded system to realize the online detection and control of the seeding performance of blanket-type tray seedling.The test results showed that the average processing time of this device for each blanket-type tray image was 4.016 seconds,and it can meet the real-time online detection requirements at a productivity of 500 trays per hour.The average accuracy of the device for detecting the seeding quantity of the rice seedling production line can reach 97.84%.When the seeding quantity of the experimental blanket-tray was set to 40 g per tray and 50 g per tray respectively,the intelligent regulation of seeding quantity can be realized.The coefficients of variation under these two kinds of seeding quantities were 2.26% and 1.60% respectively.It showed that the device can accurately detect and adjust the seeding quantity of blanket-type trays in real time,and can significantly improve the seeding quality of rice seedling production lines.(3)Using the BP neural network model,the seeding performance detection and control method research of hybrid rice pot tray seedling was carried out to achieve accurate control of the pot tray seeding quantity.A fixed threshold segmentation method was proposed,by which the seed grid image and seed image can be segmented from the pot tray image.The pixel projection method was used to obtain the pixel coordinates of the grid line from the grid image to locate the grids.The seed image was cut according to the coordinate position of the grid line,and the optimized BP neural network model was used to detect the seeding performance of the pot tray seedling.The test result showed that the detection accuracy of the passing rate,missing rate,reseeding rate,and average grain number of hybrid rice pot tray seedlings were 97.77%,95.38%,97.18%,and 97.37%,respectively.The average time to detect a single pot tray was 5.630 seconds.In order to ensure the precise seeding quantity of1-3 seeds per hole,the seeding quantity was set to 2.2 seeds per hole.The seeding quantity can be automatically adjusted near the set value according to the deviation of the seeding quantity,and the coefficient of variation of the seeding quantity was 2.87%.(4)In order to improve the detection accuracy of hybrid rice seedlings,research on the method of detecting the seeding performance of hybrid rice pot trays based on deep learning has been conducted,which has significantly improved the detection accuracy.To further explore and improve the detection accuracy of hybrid rice seeding,a U-Net network-based semantic segmentation model of grid images was established.Using this model,the grid images can be clearly and completely segmented from the pot tray images.The average pixel accuracy of the segmentation was 95.8%.Based on this,a model of hybrid rice hole-seedingquantity detection based on convolutional neural network was established to automatically extract image features and detect the hole seeding quantity.Applying the two deep learning network models with the embedded system,the detection of seeding performance of pot tray come true.The test results showed that using the Raspberry Pi system and the deep learning network model to detect the seeding passing rate,replay rate,missing rate and average grain number of the pot tray can reach 98.55%,97.18%,93.83% and 98.14 % respectively.It costs an average of 18.753 seconds to finish a single pot tray image,which can be further used for accurate testing of the seeding performance of pot tray.(5)The developed intelligent constant-quantity seeding device was installed on the rice seeding production line,forming the 2ZSB-500 Intelligent Rice Seedling Production Line,and the whole machine was tested.The performance indicators were significantly improved,and the hybrid rice intelligent,accurate and constant-quantity seeding requirements were met.The test results show that at a productivity of 500 trays per hour,using a blanket tray for seedling seeding,setting the seeding amount to 50 g per tray,the sowing amount for blanket tray seedling seeding can be maintained within the set value ±2g.The coefficient of variation of the sowing amount is 1.25%,and the sowing performance is stable and reliable.When using pot trays to raise seedlings,the seeding quantity was set to 2.2 seeds per hole.The passing rate of Wuyou 1179 hybrid rice could reach 91.82%,and the missing rate was0.53%;the passing rate of Peizataifeng seeding can reach 94.51%,and the missing rate was1.68%. |
PDF Full Text Request