[2F1-GS-10f] AI応用：安全

[2F1-GS-10f-04] A Study on Type Classification and Object Detection to be Repaired of Manhole Cover Using Deep Learnig

〇Hiroki Oikawa¹, Reon Katsumura¹, Masaki Wada¹, Hiroki Shimabara¹, Takaaki Aihara¹ (1. Nippon Telegraph and Telephone East Corporation)

Keywords:Infrastructure, Object detection, CNN, Deep learning, Inspection

Currently, we inspect annually about 60 thousand manholes within NTT East’s operation area. Technicians take pictures of manhole cover and inside manhole on site. The repair judgement of manhole equipment using these pictures is carried out visually by many people at the centralized inspection center. Last year, we created an artificial intelligence for inspecting the inside manhole using Mask-RCNN, one of the CNN algorithms, to reduce work amount of the visual check at centralized inspection center. Also we reported the accuracy of repair judgement. In this study, based on last year’s findings, we create an artificial intelligence for type classification and repair judgement of manhole cover using Mask-RCNN. In addition, we verified the detection accuracy. As a result, based on the macro average of each detection class, the recall was 96 % and the precision was 98 % for type classification, and the recall was 90 % and the precision was 70 % for repair judgement.