基于特征图的AI模型压缩方法研究及在广电行业嵌入式设备中的应用  

Research on AI Model Compression Method Based on Feature Maps and its Application in Embedded Devices in the Broadcasting and Television Industry

作　　者：张苓轩 姜竹青[1] 王海婴[1] Zhang Lingxuan;Jiang Zhuqing;Wang Haiying(Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学,北京100876

出　　处：《广播与电视技术》2024年第4期34-38,共5页Radio & TV Broadcast Engineering

基　　金：基金会项目(No.A14B07C02-202305D1)资助。

摘　　要：以深度学习为代表的AI模型在近十年中充分赋能广电行业的发展,大幅提升了内容制播、媒资管理、监测监管以及运行维护等工作的效率。本文提出一种基于特征图信息的AI模型简易压缩系统,旨在降低模型计算成本和提升模型部署效率,从而在广电行业的嵌入式设备上实现降本增效。该系统通过利用特征图信息,实现了卷积核计算成本的降低,同时保持了模型的原有性能。实验结果表明,该系统能够在基本保持原有模型能力的情况下大幅降低模型计算成本,缓解了模型部署和实时运行的压力。AArtificial Intelligence(AI)models represented by deep learning have fully empowered the development of the broadcasting industry in the past decade,significantly improving the efficiency of content production and broadcasting,media management,monitoring and supervision,and operation and maintenance.This paper proposes a simple compression system for AI models based on feature graph information,aiming at reducing model computation cost and improving model deployment efficiency.Thus,cost reduction and efficiency increase can be realized on embedded devices in the broadcasting industry.The system achieves the reduction of convolutional kernel computation cost by utilizing feature graph information,while maintaining the original performance of the model.Experimental results show that the system is able to significantly reduce the model computation cost and alleviate the pressure of model deployment and real-time operation while basically maintaining the original model capability.

关 键 词：人工智能 内容制播 监测监管 特征图 卷积神经网络 模型部署 嵌入式设备 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]