Research On Tea Bud Recognition Method In Complex Environment Based On Deep Learnin

At present,the picking of tea shoots is mainly done manually,supplemented by machinery.Manual picking requires a lot of time and labor,and now many rural youth are working in the city,resulting in no one to pick tea during the harvesting period,causing great economic losses to tea farmers.Although there are some tea bud picking machines on the market,they are not selective for the buds and are in a "one-size-fitsall" mode,which is not only very destructive to the buds,but also not conducive to the picking of famous tea.Therefore,this paper researches the deep learning-based tea bud recognition method in complex environment to lay a certain theoretical foundation for the intelligent picking of premium tea.The main research contents are as follows:(1)The denoising pre-processing and dataset construction of tea shoot images under complex environment were studied.To address the problem of noise in the captured original image and the subsequent deployment will affect the image quality,this paper proposes a new threshold function,which is used to denoise the image preprocessing.To address the problem of insufficient number of samples in the dataset,this paper adopts data enhancement methods such as rotation,flip and luminance transformation to expand the number of samples,uses the open source labeling software Label Img to label all the tea shoot images,and writes a Python program to generate the dataset by format transformation and other operations.(2)The benchmark model is selected and an improvement plan is proposed.Combining the experimental data on the homemade dataset and industrial recognition requirements,YOLOv3,which has good performance in recognition accuracy and speed,is selected as the benchmark network model instead of YOLOv4 and YOLOv5,which are not accurate for small target recognition and have relatively complex networks,and are not stable enough.problems of insufficient recognition accuracy and localization bias,and propose an improved solution to the original network model.(3)Proposing this paper’s improved model of tea shoot identification under complex environment YOLOv3-C-1.To address the problem of insufficient recognition accuracy,this paper fuses the output feature map of the second residual module with the feature map features of the last predicted output for up-sampling and introduces the spatial pyramid structure and CBAM attention mechanism in the network.To address the problem of localization bias,the K-means++ algorithm is used to re-cluster the prior frame and change the loss function to GIOU.Finally,in the experiments on the test set,the accuracy and recall of the improved model are improved by 7.8% and 10.7%,respectively,compared with the original model.