ln recent years , with the rapid development of artificial intelligence and deep learning technologies.facial recognition technology has been widely applied in aecess control systems due to its advantages ofcontaetless operation , high-precision recognition, and strong security.However,existing facial recognition accesscontrl systems still face umerous challenges in perormanee optimization , algorithm seleetion, and dataprocessing.These inelude balancing aecuracy and response speed, achieving effieient operation on resouree-constrained hardware devices , and ensuring data privacy and system reliability,'To address these issues , thisstudy , based on Huawei 's Aseend-developed MindSpore platform, conducts a comparative analysis of faciallandmarking and training algorithms, foeusing on the advantages and disadvantages of OpenCV and dlib infaeial landmarking. Experimental results indieate that dlib ofers superior accuracy and robustness inlandmarking,Therefore, dlib is used for initial facial feature point deteetion, and the ResNet deep leamingmodel is employed for deep feature extraction and classification training on multi-view facial datasets. Forhardware implementation ,the system utilizes Raspberry Pi 4B and Arduino tiNO development boards , with theML.X90614 infrared camera for real-time image acquisition and processing,The system is further enhaneed withmuli-target recognition , visitor aecess management ,and a visualized (jl design ,resulting in a high-perormancefacial recognition access control system,The findings of this study provide an effcient and reliable solution forthe intelligent upgrading of aeeess contrl systems and further validate the practical application potential ofMindSpore in small-scale embedded systems.