首页|Data from JiMei University Advance Knowledge in Intelligent Systems (Dual-student Knowledge Distillation for Visual Anomaly Detection)
Data from JiMei University Advance Knowledge in Intelligent Systems (Dual-student Knowledge Distillation for Visual Anomaly Detection)
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
By a News Reporter-Staff News Editor at Robotics & Machine Learning DailyNews Daily News - Researchers detail new data in Machine Learning - Intelligent Systems. According tonews reporting originating from Fu jian, People’s Republic of China, by NewsRx correspondents, researchstated, “An omaly detection poses a significant challenge in the industry and knowledge dist illationconstructed using a frozen teacher network and a trainable student netw ork is the prevailing approachfor detecting suspicious regions. Forward and rev erse distillation are the main ways to achieve anomalydetection.”Financial supporters for this research include Natural Science Foundation of Xia men, National NaturalScience Foundation of China (NSFC), Xiamen, China, US Depa rtment of Education, Fujian Province ofChina, Xiamen Science and Technology.Our news editors obtained a quote from the research from JiMei University, “To d esign an effectivemodel and aggregate detection results, we propose a dual-stud ent knowledge distillation (DSKD) based onforward and reverse distillation. Tak ing advantage of the priority of reverse distillation to obtain high-levelrepre sentation, we combine a skip connection and an attention module to build a rever se distillationstudent network that simultaneously focuses on high-level repres entation and low-level features. DSKDuses a forward distillation network as an auxiliary to allow the student network to preferentially obtain thequery image. For different anomaly score maps obtained by the dual-student network, we use s yntheticnoise enhancement in combination with image segmentation loss to adapti vely learn the weight scores ofindividual maps. Empirical experiments conducted on the MVTec dataset show that the proposed DSKDmethod achieves good performan ce on texture images as well as competitive results on object imagescompared wi th other state-of-the-art methods.”
FujianPeople’s Republic of ChinaAsiaIntelligent SystemsMachine LearningJiMei University
NETL
NSTL
