[2F1-GS-10f] AI応用：安全

[2F1-GS-10f-05] Differences in crack detection accuracy due to some deep neural networks

Comparison when SegNet,U-Net,FusionNet are applied to civilengineering structures

〇Masazumi Amakata¹, Ryuto Yoshida¹, Junichiro Fujii¹, Yuta Takahashi¹, Junichi Okubo¹ (1. YACHIYO ENGINEERING CO.,LTD.)

Keywords:SegNet, U-Net, FusionNet, Civil Engineering

The concrete structures are applied to various social asset facilities like roads, bridges, banks, tunnels, and sewers and support our lives. We will enter the maintenance era and the population decreasing era, and these structures are required as useful inspections and assessments as possible. Image processing is expected as a way of efficient inspections. Especially crack detection with the deep neural network is often seen as examples of Poc. But those examples are limited to each field application and don't lead us to the technical system which social assets technician, due to its nature must establish. The Deep neural network truly improved the generalization of image processing. When we apply their technologies to each industry, we must try our technical adjustments and ingenuities like network structures and parameters. Now I will focus on semantic segmentation which we often use in crack detection. I will uncover differences between SegNet, U-Net, and FusionNet, which we select as a semantic segmentation method, and refer to crack detection accuracy due to the function. As a result, I will organize a part of their knowledge leading to the civil engineering field's technology system.