AMCD : an accurate deep learning-based metallic corrosion detector for MAV-based real-time visual inspection

Yu, Leijian and Yang, Erfu and Luo, Cai and Ren, Peng (2021) AMCD : an accurate deep learning-based metallic corrosion detector for MAV-based real-time visual inspection. Journal of Ambient Intelligence and Humanized Computing. ISSN 1868-5137 (https://doi.org/10.1007/s12652-021-03580-4)

[thumbnail of Yu-etal-JAIHC-2021-AMCD-an-accurate-deep-learning-based-metallic-corrosion-detector]
Preview
Text. Filename: Yu_etal_JAIHC_2021_AMCD_an_accurate_deep_learning_based_metallic_corrosion_detector.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo

Download (9MB)| Preview

Abstract

Corrosion has been concerned as a serious safety issue for metallic facilities. Visual inspection carried out by an engineer is expensive, subjective and time-consuming. Micro Aerial Vehicles (MAVs) equipped with detection algorithms have the potential to perform safer and much more efficient visual inspection tasks than engineers. Towards corrosion detection algorithms, convolution neural networks (CNNs) have enabled the power for high accuracy metallic corrosion detection. However, these detectors are restricted by MAVs on-board capabilities. In this study, based on You Only Look Once v3-tiny (Yolov3-tiny), an accurate deep learning-based metallic corrosion detector (AMCD) is proposed for MAVs on-board metallic corrosion detection. Specifically, a backbone with depthwise separable convolution (DSConv) layers is designed to realise efficient corrosion detection. The convolutional block attention module (CBAM), three-scale object detection and focal loss are incorporated to improve the detection accuracy. Moreover, the spatial pyramid pooling (SPP) module is improved to fuse local features for further improvement of detection accuracy. A eld inspection image dataset labelled with four types of corrosions (the nubby corrosion, bar corrosion, exfoliation and fastener corrosion) is utilised for training and testing the AMCD. Test results show that the AMCD achieves 84.96% mean average precision (mAP), which outperforms other state-of-the-art detectors. Meanwhile, 20.18 frames per second (FPS) is achieved leveraging NVIDIA Jetson TX2, the most popular MAVs on-board computer, and the model size is only 6.1MB.

ORCID iDs

Yu, Leijian, Yang, Erfu ORCID logoORCID: https://orcid.org/0000-0003-1813-5950, Luo, Cai and Ren, Peng;