Title: RepVgg-Tree: a multi-branch tree based parametric reconstruction network for tiny steel surface-defect detection

Authors: Lingyun Zhu; Yuemou Jian; Lu Wang; Chongliu Jia; Xiaoxia Ma

Addresses: Liangjiang International College, Chongqing University of Technology, Banan, Chongqing, China ' Liangjiang International College, Chongqing University of Technology, Banan, Chongqing, China ' Liangjiang International College, Chongqing University of Technology, Banan, Chongqing, China; Chongqing Jialing Special Equipment Co., Ltd., Chongqing, China ' Liangjiang International College, Chongqing University of Technology, Banan, Chongqing, China ' Liangjiang International College, Chongqing University of Technology, Banan, Chongqing, China

Abstract: The surface defect detection of steel plates is an essential quality control process in digital manufacturing factories today. Traditional inspection technologies based on image processing are challenging to give a complete solution of tiny defect scale and time-consuming. We propose a novel convolutional neural network with multi-branch tree architecture, of which RepVgg-Tree can be used to realise the decoupling of training-time and inference-time architecture through structure reparameterisation. Each branch has a Vgg-like inference-time body, while the training-time model has a multi-branch topology. Furthermore, the Double-Tree attention mechanism is utilised to local feature position, improve the model's expression ability and reduce information loss. Also, the generalised focal loss is introduced to deal with the imbalance problem of positive samples and negative ones. Meanwhile, experimental results show that our model achieves satisfactory AP and average accuracy performance 95.85% and 95.96% respectively, effectively detecting tiny steel surface defects on the NEU data set.

Keywords: defect detection; RepVgg-Tree; Double-Tree; attention mechanism; tiny defect.

DOI: 10.1504/IJWMC.2023.131323

International Journal of Wireless and Mobile Computing, 2023 Vol.24 No.3/4, pp.329 - 340

Received: 10 Aug 2022
Received in revised form: 22 Nov 2022
Accepted: 12 Dec 2022

Published online: 06 Jun 2023 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article