Intelligent Treatment Of Tea Tree Pests Based On Deep Learning

Tea is widely planted in China and is an important foreign exchange earning crop.Tea tree pests are one of the main factors that harm tea plantation production and cause considerable losses to tea plantations and other related industrial chains in China every year.In order to ensure the quality of tea production is enhanced,it is important to monitor tea tree pests in a timely manner and take the correct treatment.Due to the large morphological changes of tea tree pests in the growth cycle and the lack of experience of tea plantation managers in tea tree pest identification,control measures and other related issues,this leads to the inability to accurately grasp the timing and prescribe the right medicine,unsatisfactory pest control effects and excessive use of pesticides,which directly reduces the hygiene standards and tea quality of tea produced in tea plantations.Therefore,the rapid and accurate identification of tea tree pest species and the timely adoption of scientific and effective control measures have become an urgent problem in the current tea garden industry chain.This paper focuses on the identification of tea tree pests in the following aspects.(1)This paper constructs a dataset containing 22 categories of pests named TPID.but considering the morphological differences of some tea tree pests in the growth cycle,the larvae as well as adults of some pests are divided into two categories,and the final dataset contains 29 categories of pests,including single target images of each category,and multi-target images of the same species of pests,fully considering the target detection and instance segmentation experimental process The final dataset contains 29 classes of pests,including single target images of each class and multi-target images of the same species,fully considering the situation that will occur during the target detection and instance segmentation experiment.At present,there is no dataset used for tea tree pest detection and segmentation,or related to computer vision in China,and the proposed TPID dataset will help future academic research on tea plantations.(2)For the task of tea tree pest detection and identification and instance segmentation,this paper uses the Mask Scoring R-CNN network,on top of which the attention mechanism is incorporated in the original backbone network,named MS-A R-CNN.channel attention enables the network to adaptively adjust the feature of each channel by adjusting the dependency of each channel as well as the feature The final model is able to improve the accuracy(AP50)of instance segmentation(segm)by 2.9percentage points.The pooling layer in the convolutional neural network directly merges the information,which may lead to some key information not being recognized,and spatial attention can avoid information loss,enabling the model to improve the accuracy(AP50)of target detection box(bbox)by 2.8%.(3)The design of a tea tree pest identification application for Android was completed,combining the deep learning network model and the mobile terminal.It can realize tea tree pest information query,tea tree pest information database and tea tree pest identification functions,which makes the tea tree pest identification system more intelligent and professional to assist tea plantation production.