Statistical Model Of Spike Based On Deep Learning

Rice is one of the most important sources of food for our people.In recent years,due to intensified environmental pollution,the imbalance between urban and rural population has become larger and larger,and the increase of rice production has slowed down.How to cultivate high-quality varieties of rice,tap the maximum potential of rice to resist pests and diseases,and increase rice yield per unit area has become a hot spot in current rice research.Because the phenotype is affected by environmental factors and genetic interactions,phenotype of rice is an important evaluation indicator of high-quality rice.Among them,the number of spikes and the fullness of grain are one of the most import factors for rice yield and quality.In the research of rice breeding,various phenotypic parameters including the number of spikes need to be measured at various stages of rice development,which provides a reference for subsequent breeding of high-quality varieties.The traditional method of statistics of spikes is mainly executed by manual measurement,which not only requires a large amount of labor,but also the statistical result is subjective.It is used with the disadvantages of low efficiency and possible damage to the crop growing environment.With the advancement of vision technology,the number of spikes based on image processing has become possible.However,the research in this field mainly focuses on the counting of spikes in individual plant.Due to the continuous improvement of breeding research,a statistical method for spikes that is applicable to large areas with high detection accuracy is needed to be raised to break through the Inherent bottleneck of traditional methods.In view of the above problems,this paper uses computer vision technology to propose a detection model of spikes based on deep learning with the purpose of achieving spikes counting and provide technical support for subsequent rice health monitoring and rice breeding research.The main contents include:(1)Method of counting based on spike region.On the basis of summarizing the research status at home and abroad,this paper adopts the method of extracting region of spike as the support for counting the number of spikes,and constructs a target detection model Faster R-CNN based on the ResNet101 feature extraction network to identify and locate the it’s region;(2)Explore the improvement of the model,and propose a candidate region extraction network RPN based on cluster algorithm.Based on the summary of candidate frame extraction algorithms,a clustering method named K-means for labeled box scales is proposed to bring prior knowledge to the RPN network’s generation process,further improve the recognition accuracy of the spikes,and establish a stronger detection model with more robustness..This experiment was performed in a rice experiment greenhouse in the College of Materials and Energy,South China Agricultural University,Tianhe District,Guangzhou,Guangdong Province.The rice was photographed for 2 months from heading to harvest in.In this experiment,the collected 2,000 rice ear images were used as the training set based on the Faster R-CNN detection model in this paper.The data set was manually labeled to form 1600 training sets and 400 test sets.Through testing,the accuracy of the improved detection model was 85.9%.which was increased by 1.9% between the original model especially improves the accuracy rate at the stage of heading and flowering,and our experiments show the effectiveness and rationality of the improved method proposed in this paper.Besides,by comparing our model with other one-stage models of target detection,we verify the advantages of the Faster R-CNN model adopted in this paper.Therefore,the model designed in this paper can meet the needs of rice planting scenarios and provides a data basis for subsequent rice health monitoring and optimizing breed,which has certain application value.