Research On Wheat Ear Counting Method Based On Deep Learning

To ensure national food security and formulate reasonable food prices and macro-control policies,it is necessary to predict and estimate the expected yield of crops in a timely and accurate manner.One of the critical indicators for assessing wheat yield is the number of spikes per unit area.How to accurately detect the number of spikes per unit area plays an essential guiding role in agricultural production management decisions.The traditional statistics of spikes per unit area are usually counted manually,and the results are more accurate.Still,the human resource cost is high,time-consuming,and laborious.To solve this problem,based on the target detection of deep learning,this paper conducts an in-depth study on the issue of field wheat counting.It proposes a set of field wheat ear counting systems based on deep learning,which improves the efficiency and accuracy of wheat ear counting in actual production,reduces labor costs,and provides a reference for automation and intelligent management of actual output.The specific contents of this article are as follows:1,Firstly,the wheat image data set was established by collecting wheat images through the GWHD website and shooting wheat ear images in the field.Then,image preprocessing is performed through image processing techniques such as image segmentation,image annotation,and data enhancement to prevent problems such as data imbalance and image noise from affecting model training.Finally,Labeling image annotation software is used to annotate the data.After the annotation,the image data is made into the VOC2012 data set format for model training.2.According to the difference between the single and double stages of target detection,based on the Pytorch deep learning framework,the YOLOX network model and the Faster RCNN network model are built,and the two networks are optimized and adjusted for the wheat ear detection task.The attention mechanism is introduced into the backbone of the YOLOX network model to improve the network’s attention to dense wheat ears,solve the problems of small wheat ear shape and complex feature extraction,improve the SPP part of the backbone network of the model,and optimize the model structure to improve the network training speed and prediction accuracy.Adjust the strategy of the prediction stage and use Soft-NMS to predict overlapping targets.3.In the Faster R-CNN target detection algorithm,the Anchor box of the model is modified utilizing clustering,which significantly alleviates the problem of low detection accuracy caused by the significant difference in size between the genuine bounding box and the model prediction bounding box.And the transfer learning technique is used in the wheat ears detection task to accelerate the model fitting speed and improve the model’s generalization simultaneously by the training method of freezing first and then thawing.From the experimental results,the improved models performed in this paper are better than the initial models.4.The wheat ears detection system is designed according to the improved network model,which mainly has two functions of image detection and dynamic video detection,and realizes the function of uploading the pictures of wheat ears in the field to derive the counting results in real-time,which helps different people to learn and study this aspect and provides better user experience and convenience for wheat production management.