Research On Identification Of Tea Cultivars Based On Deep Learning

Tea germplasm resources are important strategic resources in China.Accurately distinguishing different tea cultivars is the premise for the research and protection of tea germplasm resources,which will help promote the high-quality development of my country’s tea industry.In this paper,we first built a tea leaf image dataset,and then used a variety of deep learning models to establish some tea cultivars identification models,then selected the better model through performance comparison,and then obtained the optimal identification model through candidate frame optimization and designed an experimental analysis model generalization ability,and finally complete the development of tea cultivars identification software based on the optimization model.The main research work and results completed are summarized as follows:(1)According to the research needs,select 7 kinds of tea cultivars: Anji Baicha,Baihaozao,Nongkangzao,Shuchazao,Wuniuzao,Longjing 43 and Huangkui,and design the leaf selection and picking strategy,self-built tea leaves(real leaves)dataset.(2)Combined with literature research and experimental conditions,Faster R-CNN,SSD,YOLOv3 and YOLOv4 were selected as candidate identification models to compare the identification performance of each model.The experimental results show that the identification performance of the YOLOv4 model is better than the other 3models.Optimize the YOLOv4 model.The candidate frame of the YOLOv4 model is optimized based on the K-Means++ algorithm.The experimental results show that the identification accuracy of the optimized YOLOv4-i model is significantly improved.(3)Experiments and discussions were carried out on possible factors affecting the accuracy of the model,The results showed that the YOLOv4-i model showed good identification ability for the identification of more cultivars of tea and under different environmental brightness,but the identification of tea under different growth environments may need to be improved.The experimental results show that the accuracy rate of the YOLOv4-i model in this paper is 96.43% in the validation set and 98.776% in the self-made test set and it shows better variety generalization ability.