Research on AI Model Compression Method Based on Feature Maps and its Application in Embedded Devices in the Broadcasting and Television Industry
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.
Artificial intelligenceContent production and broadcastingMonitoring and supervisionFeature graphConvolutional neural networkModel deploymentEmbedded device
NETL
NSTL
万方数据
