Unstructured Multi-view Stereo Anomaly Detection
Konferenz: ICMLCA 2021 - 2nd International Conference on Machine Learning and Computer Application
17.12.2021 - 19.12.2021 in Shenyang, China
Tagungsband: ICMLCA 2021
Seiten: 4Sprache: EnglischTyp: PDF
Persönliche VDE-Mitglieder erhalten auf diesen Artikel 10% Rabatt
Autoren:
Lu, Xudong; Bai, Fan (Equipment Engineering College, Shenyang Ligong University, Shenyang, China)
Li, Lun (Information and Control Engineering College, Weifang University, Weifang, Shandong, China)
Zhang, Hui (Mechanical Engineering College, Shenyang Ligong University, Shenyang, China)
Inhalt:
With the development of machine vision, product defect detection based on image recognition is more and more widely used in industrial production, which greatly improves the production efficiency of products. However, the traditional detection method based on two-dimensional image is limited by the shooting angle and can not effectively distinguish the defects from the perspective of space. The unstructured multi-view stereo anomaly detection (UMVSAD) proposed in this paper obtains the depth map by fusing the multi angle photo information through the deep learning network, and uses the latent feature to detect whether the product has defects. Experiments show that the average AUC of the generated data of DTU dataset is higher than 0.578. For self-built datasets, the average AUC is higher than 0.721, lay the foundation for follow-up work.