Research On Detection Method Of Wheat Ear In Field Based On Image Deep Learning

Wheat as one of the main grain crops in China,timely and accurate information on wheat production,for crop production management,to ensure the normal distribution of grain and price regulation,to ensure national food security is of great significance.In agricultural production activities,for wheat,a crop with variable number of tillers,ear density is one of the key agronomic indicators of wheat yield.In the past,the number of wheat ears per unit area in a large field environment was mainly obtained by manual counting,which consumed a lot of human and material resources,was less efficient and was affected by subjective factors.With the development of artificial intelligence technology,deep learningbased object detection technology is widely used in agriculture.In the task of wheat ear detection and recognition,more and more researchers are using easily accessible RGB images of wheat ears to carry out research on wheat ear detection and recognition to accurately and quickly.Thus,wheat production is predicted.In this paper,we take wheat ears images in a large field environment as the research object,and we aim to improve the precision of wheat ear detection model and reduce the workload of labelling wheat ears manually.Constructing YOLOv5-ear wheat ear detection model and proposing Efficient YOLOv5-ear wheat ear detection method based on semi-supervised learning.Provide technical support for automatic detection and counting of wheat ear in a field environment.In turn,it provides a reference basis for wheat yield prediction,wheat breeding and variety selection.the main research contents include:(1)A study of YOLOv5-ear wheat ear detection method based on supervised learning.Faced with the problems of small wheat spike size,dense distribution,complex growth background,and susceptibility to wheat ear and leaf shading in a large field environment.To improve the performance of the wheat ear detection algorithm,from the perspective of enhancing the learning of wheat ear feature information.Using YOLOv5 s as the baseline model,Adding the SENet Attention Module to the backbone.Adding the CBAM Attention Module to the Neck Network.Replacement of the upsampling method in the neck network for CARAFE.K-means clustering anchor box for the wheat ear dataset,balancing the different scale confidence loss values with the real anchor frame distribution.The final YOLOv5-ear wheat ear detection model was constructed.Model performance evaluation using the MWD,the self-built wheat ear dataset.The YOLOv5-ear model achieved 93.7%,93.3%,93.5% and 94.9% for the precision P,recall R,harmonic mean F1 and mean Average Precision m AP,Compared to YOLOv5 s added 1.6%,1.7%,1.6% and 1.8%,respectively.The coefficient of determination R2 between the YOLOv5-ear model detection counts and manual statistics was 0.95,the RMSE was 4.48,and the MAE was 3.78.Model performance evaluation using the GWD,the global wheat detection dataset,the m AP of the YOLOv5-ear model is 94.8%.Experimental results show that the YOLOv5-ear model can improve the detection performance of dense wheat ear in a large field environment(2)A study of Efficient YOLOv5-ear wheat ear detection method based on semisupervised learning.Currently,most of the ear detection models are trained by supervised learning.Dependence on labeled data and its quantity.However,the process of producing labeled data is very time-consuming and labor-intensive.In order to reduce the workload of manual labeling of wheat ears,wheat ear detection is performed using small-scale labeled wheat ear data and a large amount of unlabeled wheat ear image data.Combining the semisupervised learning approach based on the Efficient Teacher’s semi-supervised object detection model.Using the teacher-student mutual learning model framework.Improving the method of identifying reliable pseudo labels.Using the YOLOv5-ear wheat ear detection model as the wheat ear object detector.Construction of a semi-supervised learning based Efficient YOLOv5-ear wheat ear detection model.Model performance evaluation using 20% of the MWD labeled wheat ear data.The m AP of the Efficient YOLOv5-ear model reached 95.3%.It improves by 0.4% compared to the YOLOv5-ear wheat ear detection model which is using all wheat ear data.The coefficient of determination R2 between the Efficient YOLOv5-ear model detection counts and manual statistics was 0.95,the RMSE was 4.34,and the MAE was 3.70.Using 20% of the GWD labeled wheat ear data for model performance evaluation,the m AP of the Efficient YOLOv5-ear model reached 95.2%.The experimental results show that the Efficient YOLOv5-ear wheat ear detection model trained using smallscale labeled data.Compared to the YOLOv5-ear wheat ear detection model trained using all labeled data.It can achieve the same level of detection accuracy and counting effect.