A Crop Row Extraction Technology For Paddy Rice Seedlings Based On Convolutional Neural Networks

China is a primary rice planting country,and its rice plant area is about18.4% of the world’s.The production of grain in 2018 was up to 27.6% of the world’s which ranking the first place in the world.However,the irregular distribution of rice seedlings in paddy field caused by the faultily artificial driving transplanter in the deep and inhomogeneous paddy soil make it difficult to apply mechanical management devices for the serious seedlings’ destruction and high management cost.In order to liberate the mechanical management devices application and reducing the management cost,we developed an automatic driving system based on computer vision for rice seedling transplanter navigation.In this paper,the crop rows detection method based on convolutional neural networks was designed so that the transplanter would drive following the target row center line during transplanting.In this way,the rice seedlings distribution in paddy filed would be regulate.Furtherly,an embedded crop row detection system was built up in circuit board for transplanter’s computer vision navigation system.The research contents and conclusions are as follows:(1)Constructing rice seedlings image databaseComponents were choiced according to the requirement of transplanter navigation system.And the camera installation mechanism was built up to reduce vibration from tranplanter.This research designed a device to acquire images by analyzed the positional relationship of transplanter,camera,rice seedlings and integrating with imaging principle of camera.A large number of rice seedings in sample images acquired in different timing within one day,illumination intensity and angle were artificial labelled out in the way of stem-based as rice seedling training datasets.(2)Build up an deep-learning model for rice seedling detection and locationThose datasets were used to train a deep learning model called Faster R-CNN with 3different feature extraction networks(VGG network,VGG＿CNN＿M＿1024 network,ZF network)in 2 training method(approximate joint training method and alternating training method).The m AP(mean average precision)and detecting rate was calculate the trained model’s performace in rice seedling detection and location.The result showed that the model with VGG＿CNN＿M＿1024 network trained with approximate joint training method has considerable precision with 86.9% and real-time perfomance with 0.81.And the highest m AP which got 89.9% was the batch size of 256 and learning rate of 0.005.(3)A crop row detection algorithm for paddy rice seedlings based on Faster R-CNNIn this paper,Images were fed into a trained convolutional neural network,Faster R-CNN and the outputs were the locations of seedlings.Seedlings in the reference seedling row for transplanter navigation were extracted with the agglomerative hierarchical clustering method,and the centers of the seedlings were fitted with the linear least square method to form the center line of the reference row.Results showed that the proposed algorithm produced low errors,in terms of angle error(deviation)and initial point error(side shift).The average of the angle error was 0.60° and RMSE was 0.07°in a range of0.45-0.67° over all the six seedling spacings tested.The average of the initial point error was 4.75 pixels(equivalent to 10.39 mm)with RMSE approximately 1.16 pixels(2.54mm)in the range between 3.30-6.58 pixels(7.22-14.39 mm).Thus,the proposed algorithm could provide reasonably accurate navigation for seedling transplanting in paddy fields and showed strong promise of developing real-time location navigation system for rice planters.(4)An embedded rice seedlings rows extraction method in Jetson TX2In this paper,we analyzed embedded hardware platform and calculation accelerator,we used Jetson TX2 embedded board,which is based on GPU accelerator,as an embedded platform to develop a seedling belt line extraction algorithm based on deep neural network modelling.And built an embedded operating system Linux for Tegra in Jetson TX2.In terms of visual requirement,we installed OpenCV,VisionWorks computer vision and image processing tools;we also installed CUDA,Cu DNN for convolutional neural network model to accelerate the operation of GPU.Furthermore,we transplanted the algorithm and tested the arithmetic speed.The result shows that the average operating rate is 77 ms.The embedded computer vision system’s real-time performance was 0.879.