SoC 기반의 엣지 디바이스를 위한 Yolov5M 경량화 및 최적화에 관한 연구

Recently, intelligent systems with artificial intelligence have been basically adopted in the field of CCTV, and the method of analyzing images input from CCTV using deep learning algorithms is mainly used in the server. However, image quality deterioration may occur in the process of transmitting and storing images, reliability is difficult to be guaranteed in terms of security, and continuous investment in server construction and management is required. When building a CCTV network using an SoC-based edge device equipped with NPU functions, these shortcomings can be overcome and better quality services can be provided. However, since lower performance NPU must be utilized compared to the server, various studies should be accompanied to reduce the weight of the deep learning model and improve accuracy. In this paper, we implemented a deep learning model that can guarantee high speed and accuracy by optimizing and lightweighting the Yolov5M model specialized for the NPU envi ronment using the NAS (Neural Architecture Search) technique.