Title: Research on micro defect recognition based on deep learning

Authors: Donghong Yang; Yude He

Addresses: School of Mathematics Department, Xi'an Jiaotong University City College, Xi'an, China ' Data Analysis Center, Big Data Center of State Grid Corporation of China, Beijing, China

Abstract: To improve the overall recognition accuracy for micro defects, magnetic tile defect images are taken as the research object, and a defect image recognition method based on deep learning is proposed. Using MobileNetV3 network as the basic model, the number of deep convolution and the number of channels are trimmed. Then, mish function is used to replace h-swish function as the activation function, and the training speed and recognition accuracy of the model are improved, thus realising efficient and accurate recognition of micro defect images. The simulation results show that compared with the recognition methods standard MobileNetV3 network and other classification models faster R-CNN and EfficientNet, the proposed method performs better in terms of accuracy and F1 value, reaching 99.59% and 98.75%, respectively. The proposed method can recognise the defect image more accurately, and has the advantages of high inference speed, low parameter quantity and low computational cost.

Keywords: image recognition; image segmentation; deep learning; MobileNetV3 network.

DOI: 10.1504/IJBIC.2024.139260

International Journal of Bio-Inspired Computation, 2024 Vol.23 No.4, pp.257 - 265

Received: 24 May 2023
Accepted: 18 Dec 2023
Published online: 28 Jun 2024 *

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