At SMART2023 on 4th July 2023, Theodoros Tziolas, PhD student at University of Thessaly, presented the paper CenterNet-based models for the detection of defects in an industrial antenna assembly process.
Authors: Theodosios Theodosiou, Theodoros Tziolas, Konstantinos Papageorgiou, Aikaterini Rapti, Elpiniki Papageorgiou, Sebastian Pantoja, Paschalis Charalampous, Nikolaos Dimitriou, Dimitrios Tzovaras, A Cuinas, J Mourelle, Andreas Böttinger, George Margetis
Brief description: This study aimed to identify incorrect antenna assembly using Artificial Intelligence. The anchor-free and lightweight object detection CenterNets were combined with different feature extractors, and their performance was assessed in terms of the trade-off between high accuracy and low latency, as applied on a real industrial dataset with limited images of defective samples. To enhance the training dataset, synthesised defects were induced into the dataset, followed by heavy data augmentation which was performed on-the-fly during training.
Findings: The proposed approach is more effective than the binary classification approach. The CenterNet-ResNet50 architecture performed better than other state-of-the-art computer vision classifiers, as it produced the highest detection metrics and low inference speeds. Teaching the model what it must “see” or what the defect is, increased the accuracy and exploited better the dataset that contained images with more than one defect.
Future work: Future work will focus on more comprehensive datasets and more lightweight models.
Interested in finding out more? The paper will be published as an open access publication in the SMART 2023 conference proceedings.