Design And Implementation Of Online Dataset Annotation System Based On Deep Learning

In recent years,deep learning target detection algorithms have made significant breakthroughs and have been gradually applied to many fields such as intelligent security,unmanned vehicles,and medical diagnosis.These mature algorithm models are usually based on a large number of annotated images,however,the large amount of image annotation work is time-consuming and difficult to complete in a short period of time,and how to improve the efficiency of image annotation has been a hot research topic in the field of target detection in recent years.In addition,operations such as data enhancement,model training and model evaluation required for target detection applications need to be completed using different tools,which is not conducive to the rapid implementation of deep algorithm models.Therefore,this thesis investigates and analyses the popular annotation systems and develops a multi-functional and efficient dataset annotation system based on the Label Studio platform.The main work contents are as follows.1)Learning the development techniques used in building the Label Studio platform and delving into the platform’s front and back-end interaction logic,database design,front-end page building process,image upload and annotation management cycle.2)Aiming at the problems that the automatic annotation algorithm integrated in the current annotation platform is old and the automatic annotation of objects with smaller area share in the image is not satisfactory,an automatic annotation algorithm based on the improved YOLOv5 is proposed.A more efficient Bi FPN structure is used in the feature fusion part to achieve multi-scale fusion of features;a custom Dc-res-SA module is added at the small target detection end to refine the semantic information of small targets using the SA attention mechanism and some residual structures.The experimental results show that the average accuracy of the improved algorithm in this thesis is improved by 2.8%,the average accuracy value of small targets is improved by 1.2%,and the detection speed of a single image is 21 ms,which effectively improves the accuracy of the algorithm annotation.3)Completed the development and testing of the dataset annotation system.According to the requirement analysis,the system was divided into four functional modules,including data enhancement,automatic annotation,algorithm module training,training result visualization,and the internal processes and data interaction interfaces of each module in the system were designed and implemented in detail.The test results show that the system not only can effectively improve the annotation efficiency of datasets,but also support the full cycle management of YOLOv5 series algorithm training without the need of other tools,which has high practicality.