| Nowadays,tea plays a very important role in daily life and it can be said that one cannot live without tea.The demand for tea production and deep processing is also constantly increasing.Tea is abundant in China and has a wide planting area,especially in Yunnan Province,where the tea planting area ranks first and the output value second in China.The grade of fresh tea leaves is an important factor affecting the output value of tea.However,at present,the recognition of machine-harvested fresh tea leaves relies on manual methods with which there are problems such as poor recognition effect,low efficiency,and strong subjectivity,severely limiting the improvement of economic benefits from tea.Therefore,developing an efficient and rapid(automatic)recognition technology for machine-harvested fresh tea leaves can not only improve the economic benefits to tea farmers but also provide academic reference for other related researches.In this paper,machine-harvested large-leaf Pu’er tea was taken as the research object,and recognition of fresh tea leaves was studied.A method combining machine vision technology with deep learning algorithm is proposed,and its effectiveness in identifying machine-harvested fresh tea leaves is verified through experiments.More than one thousand images of machine-harvested fresh tea leaves were used for training and testing,and different shapes,colors and sizes of fresh tea leaves were identified.The results showed that the proposed method can complete the task of fresh tea leaves classification with relative accuracy.In addition,the optimization space of the algorithm was analyzed and the possible improvement scheme was proposed in order to provide theoretical and technical reference for future research on automatic classification and quality detection of fresh tea leaves by machine.The main research work of this paper includes:(1)Establishing an image acquisition system for fresh tea leaves.The influence of hardware equipment such as light sources,industrial cameras,and lenses of the machine vision system on image quality was studied.According to the physical characteristics and requirements of fresh tea leaf picking by machine,relevant equipment of machine vision system suitable for this experiment was selected to build an image acquisition system of fresh tea leaves,and complete relative data and image collection.(2)Creating a dataset of fresh tea leaves.During the process of collecting fresh tea leaf data,the data may be affected by information noises such as light sources and electromagnetic interference.In order to obtain high-quality images,the collected data were preprocessed by regional clipping,size normalization,image smoothing,image sharpening,etc.,to enhance the feature information of fresh tea leaf image.In this experiment,a median filter with a 3×3 template and a first-order Sobel operator were selected to eliminate the information affecting image features.After tone transformation,image rotation,background transformation and other data enhancement operations of the data set,comparison and analysis of the dataset were carried out to complete the construction of machine-picked fresh leaf image data set.(3)Applying machine vision technology in the field of machine-harvested tea leaf recognition.The finished dataset of fresh tea leaves were trained and compared using object detection algorithms,YOLOv5 and Faster R-CNN,to determine the best model for this experiment,which was YOLOv5 for its capability of detecting and recognizing the categories of fresh tea leaves.To address issues such as poor recognition effects and false detections in the identification,an optimized YOLOv5-based classification method was proposed.Three optimizations were made: firstly,a Bi FPN feature network was introduced to perform more feature fusion without additional cost;secondly,Io U in NMS was replaced with DIo U to solve problems such as overlapping predicted boxes;thirdly,SGD stochastic gradient descent optimizer was used to achieve better results with less computation.We experimented with the above network models on the dataset of machine-harvested fresh tea leaves and analyzed and compared optimization-related parameters.Through comparative experiments,the optimized YOLOv5 model most suitable for the dataset was selected and its network was trained.The experiment showed that the m AP of the optimized YOLOv5 model on the fresh tea leaf dataset reached 72.50%,which was 17.40%higher than that of the original YOLOv5 model.The experiment basically achieved expected result and research objectives. |
