The Research On Wheat Ear Detection Method In Field Based On Deep Learning

Wheat is one of the three major food crops in China,and its production status has a direct impact on Chinese food security.Wheat ear detection can provide an important reference for agricultural production management,rapid and accurate detection of wheat ears has important influence and significance in evaluating wheat yield.Wheat grows in a complex field scene,there are usually great differences in the size and posture of wheat ears,and there are serious overlap between ears and ears.These complex factors led to the accuracy of object detection model is low.In addition,the rapid detection in the field has higher requirements on the model,which can not only achieve rapid and accurate wheat ear detection,but also need to meet the lightweight requirements of the model,so that the detection model can be implanted into mobile devices which was conveniently used to detect wheat ears in the field.Therefore,this research aimed to improve the accuracy and speed of detection,and the wheat ear detection methods were carried out based on the deep learning.The main research results are as follows:(1)The preprocessing method of wheat ear image was studied.Aiming at the problems of blur and low resolution in some images in the wheat ear data set,image reconstruction research was carried out.Firstly,image reconstruction models based on SRCNN,NSFSRCNN,and ESRGAN networks were constructed.Then,different reconstruction models were compared,and it was found that the reconstruction model based on the ESRGAN network has the best effect.Finally,the blurred image was processed based on the ESRGAN network to generate a clear,high-resolution image of wheat ears.(2)Wheat ears image data set was constructed.Through manual collection of wheat images in the field,a deep learning-based field wheat ear image data set was constructed.In order to overcome the influence of various environmental interference factors,firstly,the original image was amplified by brightness transformation,contrast enhancement,random rotation and other data amplification processes.Secondly,the wheat spike target was manually labeled in the self-built image data set.Finally,the self-built data set was divided according to the Pascal VOC data set format and randomly divided into training set,verification set and test set.(3)A field wheat ear object detection model based on deep learning was constructed.Firstly,the wheat ear detection models based on Faster RCNN,R-FCN,SSD,and YOLOv3 were constructed respectively.The wheat ear detection test was performed on the reconstructed image and the original image using the constructed detection model.Then,the evaluation indexes of the above object detection models were compared under the same test set,and it was found that the detection accuracy of the YOLOv3 model was the highest.Then,in order to realize the lightweight model of wheat ear detection,an accelerated and reduced version of the YOLOv3-tiny model was constructed,and a field wheat ear detection model based on the YOLOv3-Efficient Net network was proposed.Finally,experiments were conducted with the self-built data set.The results showed that the mean average precision(m AP)of wheat ear detection based on the YOLOv3-Efficient Net model reached 90.91%,which was 3.65% higher than that of YOLOv3-tiny model.Compared with the test results based on Faster RCNN,R-FCN and SSD,the m AP increased by 7.53%,6.29% and 12.54%respectively,and the model size has the lowest size at 18.9MB.In this study,the proposed model not only had obvious advantages in detection accuracy and detection speed,but also had a smaller storage space,which achieved light weight so that it could be migrated to small mobile devices.