Research On Detection And Recognition Of Tea Sprouts Under Natural Environment

Tea industry is an important aspect of China’s import and export trade commodities.Tea has a long cultural heritage in China,and is closely related to the life of our people.In our country,there are mainly two ways to harvest tea sprouts,manual picking and mechanical picking.The mechanical picking is efficient,yet the different levels of tea sprouts are often mixed,making the subsequent processing screening difficult.Manual picking is more precise,however,it’s also inefficient and labor-intensive.Considering the above factors,it is urgent to carry out the intelligent picking research of tea tender sprouts.In the natural environment,the growth environment background of tea sprouts is complex,and their growth patterns are also different.Therefore,the primary task in the intelligent tea sprouts picking is detection and recognition.This paper investigates the detection and recognition technology of tea sprouts under natural environment,and combines the relevant methods based on deep learning to achieve higher accuracy of tea sprouts detection and recognition tasks.The main work of this paper is as follows:1.For the task of object detection of tea sprouts in the natural environment,this paper took the tea sprouts in tea gardens as the research object,establishing the dataset of tea sprouts under natural environment.To solve the problem that the noise present in the original tea sprouts images may affect the network input,the images were pre-processed for denoising.And about the problem of small number of samples in our homemade dataset,this paper used data enhancement methods such as flipping,up-and-down alignment,rotation,contrast adjustment,etc.to expand the number of samples in the dataset.The open source software Labelimg was used to perform the labeling work on the tea sprouts in the dataset,and then using the Python program to convert the label files into a dataset format suitable for the object detection task.2.Performed baseline model selection and proposed solutions for model optimization.Based on the experimental data results of the homemade tea sprouts dataset and the comparison of various object detection network models,the YOLOX model,which is better in terms of model novelty,recognition accuracy and speed,was selected as the basic network model on a comprehensive basis.After analyzing some problems of the original YOLOX network such as insufficient detection accuracy and object localization bias,a series of improvements were made to the original YOLOX network.3.Proposed an improved YOLOX-ST network for tea sprouts recognition and detection in the natural environment.This paper replaced Swin Transformer module as the backbone of the whole YOLOX network and added the CBAM mechanism in the part of model for multi-layer feature extraction to enhance the extraction of tea sprouts features by the network.To solve the lack of accuracy of the network for small object detection of tea sprouts,this paper added a small object layer to the model and modified the IoU loss function of the original model to an EIoU loss function with the confidence degree dominated by Focal Loss.Finally,according to the experiment results,the accuracy and recall of the YOLOX-ST model proposed in this paper were improved by5.17% and 4.81%,respectively,compared with the original model.