Semitela Ângela, Pereira Miguel, Completo António, Lau Nuno, Santos José P
Centre of Mechanical Technology and Automation (TEMA), Department of Mechanical Engineering, University of Aveiro, 3810-193 Aveiro, Portugal.
Intelligent Systems Associate Laboratory (LASI), 4800-058 Guimarães, Portugal.
Sensors (Basel). 2025 Jan 17;25(2):527. doi: 10.3390/s25020527.
To automate the quality control of painted surfaces of heating devices, an automatic defect detection and classification system was developed by combining deflectometry and bright light-based illumination on the image acquisition, deep learning models for the classification of non-defective (OK) and defective (NOK) surfaces that fused dual-modal information at the decision level, and an online network for information dispatching and visualization. Three decision-making algorithms were tested for implementation: a new model built and trained from scratch and transfer learning of pre-trained networks (ResNet-50 and Inception V3). The results revealed that the two illumination modes employed widened the type of defects that could be identified with this system, while maintaining its lower computational complexity by performing multi-modal fusion at the decision level. Furthermore, the pre-trained networks achieved higher accuracies on defect classification compared to the self-built network, with ResNet-50 displaying higher accuracy. The inspection system consistently obtained fast and accurate surface classifications because it imposed OK classification on models trained with images from both illumination modes. The obtained surface information was then successfully sent to a server to be forwarded to a graphical user interface for visualization. The developed system showed considerable robustness, demonstrating its potential as an efficient tool for industrial quality control.
为实现加热设备涂漆表面质量控制的自动化，通过在图像采集时结合偏折测量法和基于强光的照明、在决策层面融合双模态信息的非缺陷（良品）和缺陷（不良品）表面分类深度学习模型以及用于信息调度和可视化的在线网络，开发了一种自动缺陷检测与分类系统。测试了三种用于实施的决策算法：一个从头构建和训练的新模型以及预训练网络（ResNet-50和Inception V3）的迁移学习。结果表明，所采用的两种照明模式拓宽了该系统可识别的缺陷类型，同时通过在决策层面执行多模态融合保持了较低的计算复杂度。此外，与自建网络相比，预训练网络在缺陷分类上取得了更高的准确率，其中ResNet-50表现出更高的准确性。该检测系统始终能快速准确地进行表面分类，因为它对用两种照明模式的图像训练的模型都进行良品分类。然后将获得的表面信息成功发送到服务器，以便转发到图形用户界面进行可视化。所开发的系统显示出相当的稳健性，证明了其作为工业质量控制高效工具的潜力。
