Research On Apple Fruit Detection Technology Based On Data Augmentation And Transformer

The increasingly aging population leads to the serious shortage of the rural labor forces and the dramatic increasing of the labor costs,which are unfavourable for the large-scale production of apples.However,the development of new technologies,such as AI,Io T and Big Data,is helpful for changing this phenomenon.The intelligence of the modern Apple production has thus become a necessary trend.Automatic apple-picking technology is capable of effectively reducing the labor forces employed in harvesting apples.However,the key of automatic apple-picking technology lies in the identification and location of apple fruits in natural environments.So,this thesis studies an apple fruit-identificating approach based on the data enhancement and the Transformer,and develops a prototype system for this purpose that solves the core problem in designing a picking robot.The main work is as follows:(1)An apple fruit identification approach based on the data enhancement.The data enhancement is about making limited amount of data be able to generate more equivalent data.Several common approaches on the data enhancement are analyzed,which focuses on the Rand Augment,an image data enhancement approach.Then,the researching result has been applied for identifying apple fruits.The experimental results show that the Rand Augment can get a higher AP value when compared with traditional geometric data enhancement approaches.(2)An apple identification model based on the Transformer.After analyzing the CNN-based object detection model,this thesis utilizes the Transformer model in the NLP,which is good at capturing the long-range dependence of sequences,to propose a Transformer-based object identification model with a CNN as the bone for extracting feature.This model has implemented the end-to-end object identification and its accuracy is superior to traditional CNNs.(3)An apple fruit identification system is designed and implemented.On the basis of key technologies,such as the dataset construction,the image enhancement model and the apple fruit image identification model,a prototype system for identifying apple fruits and predicating the outputs is provided.