Design And Implementation Of Zero-code Development Platform For Target Detection

In recent years,the widespread application of artificial intelligence technology has had a broad and profound impact on production and life,public services,social governance,and even the global competitive landscape.As one of the important fields of artificial intelligence,computer vision based on deep learning has attracted many Attention of developers and development agencies.Among them,the demand for target detection tasks based on deep learning has increased sharply,such as firework detection,helmet recognition,etc.However,the corresponding target detection services can only be provided by developers or development organizations who understand artificial intelligence,and developers are required to understand target detection.Relevant theoretical models have a certain programming and mathematical foundation,but cannot meet the increasingly diversified and liberalized needs of enterprises.Therefore,there is a surge in development demand for zero-code development platforms for target detection that aim to lower development thresholds and improve development efficiency.This paper aims at the pain points of the existing low-code or zero-code development platform for deep learning that do not support local deployment and lack of privacy protection for training data,single data upload function,no real-time information statistics,and lack of resource management.The zero-code development platform supports one-stop zero-code development services from data management,online labeling,model management,model training to result display,and model deployment.The main work and achievements of this paper are as follows:(1)Based on domain-driven design,a zero-code development platform for target detection based on micro-service architecture is realized.Based on the idea of domain-driven design,this paper completes the reasonable division of platform sub-domains and the mapping of domain models to micro-service architectures through strategic design,realizes the decoupling between various modules of the system,and realizes the design and implementation of specific functions within the domain through tactical design.(2)Design and implement a zero-code modeling method for object detection,and improve the deficiencies of existing platforms.Based on the development steps of the deep learning task,the development process of the visual target detection task is defined,and the development of the target detection model can be completed by simply clicking on the Web visualization interface.In the zero-code modeling of the whole process,a multi-concurrent file transmission scheme based on hash value verification is designed and implemented in the data management module,which realizes the breakpoint resume and second transmission functions of large file uploads,improves the transmission rate,and solves the problem Large file upload failures cannot be continued and the same file can be uploaded repeatedly;in the online labeling module.a dynamic data allocation scheme based on data blocks is designed and implemented,which realizes the dynamic allocation of labeling data and solves the problem of low efficiency of team labeling;The model application module designed and realized the camera configuration integration and connection reasoning function with the model,and output the reasoning screen in real time.which greatly shortened the period for the model to be put into use:it supports local deployment of the system,and achieved the goal under the premise of ensuring the privacy of user training data The whole process of detection task development is zero code.(3)A solution for job lifecycle management and resource management scheduling based on OpenPBS cluster is proposed.Based on the OpenPBS job scheduling and cluster management system,functions such as job opening,job termination,job resource occupation,and cluster resource query are designed and implemented,providing support for the job operation of the model management module and the training environment configuration of the model application module,solving the traditional problem Lack of effective management of operations and resources in the program.(4)A real-time information statistics solution based on Kafka and Spark Streaming is proposed.Based on the distributed message queuing system Kafka and the stream processing engine Spark Streaming,it designs and realizes the statistics of parameters such as accuracy rate and loss function generated during the model training process,and supports drawing graphs and pie charts to display real-time training results.Various real-time parameters provide support for viewing and training real-time information of the model management module,and solve the problems of-high delay,poor scalability,and low reliability in traditional information transmission.After system testing and comparison with other existing platforms,it can be seen that the zero-code development platform for target detection proposed in this paper effectively solves the problem that the existing low-code or zero-code development platforms based on deep learning do not support training data privacy protection.Pain points such as single data upload method,no real-time information statistics,and lack of resource management are more reliable.stable,flexible,and time-sensitive.