TY - GEN
T1 - GH-QFL
T2 - 32nd International Conference on Artificial Neural Networks, ICANN 2023
AU - Xiao, Xianjing
AU - Du, Yan
AU - Yang, Rui
AU - Hu, Runze
AU - Li, Xiu
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.
PY - 2023
Y1 - 2023
N2 - In the manufacturing sector, industrial defect detection technology has become a crucial component for substantial improvements in both product quality and production efficiency. However, the accuracy of deep learning-based defect detection methods can be compromised by uneven training data, which could result in a bias towards over-represented classes. To address this issue, some hard example mining (HEM) methods have been developed to balance the contribution of different classes during training. Nonetheless, on the custom dataset, these methods still inherit the hyper-parameters predefined on the COCO dataset. We thereby propose a novel loss function, called Gradient Harmonized Quality Focal Loss (GH-QFL), to weight hard examples dynamically based on gradient statistics. The proposed approach is evaluated on a defect detection dataset: NEU-DET. The results demonstrate that our method outperforms the detection method using other loss functions by 3.1 % mean average precision (mAP).
AB - In the manufacturing sector, industrial defect detection technology has become a crucial component for substantial improvements in both product quality and production efficiency. However, the accuracy of deep learning-based defect detection methods can be compromised by uneven training data, which could result in a bias towards over-represented classes. To address this issue, some hard example mining (HEM) methods have been developed to balance the contribution of different classes during training. Nonetheless, on the custom dataset, these methods still inherit the hyper-parameters predefined on the COCO dataset. We thereby propose a novel loss function, called Gradient Harmonized Quality Focal Loss (GH-QFL), to weight hard examples dynamically based on gradient statistics. The proposed approach is evaluated on a defect detection dataset: NEU-DET. The results demonstrate that our method outperforms the detection method using other loss functions by 3.1 % mean average precision (mAP).
KW - Defect detection
KW - Hard example mining
KW - Loss function
UR - https://www.scopus.com/pages/publications/85174573421
U2 - 10.1007/978-3-031-44207-0_20
DO - 10.1007/978-3-031-44207-0_20
M3 - Conference contribution
AN - SCOPUS:85174573421
SN - 9783031442063
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 232
EP - 243
BT - Artificial Neural Networks and Machine Learning – ICANN 2023 - 32nd International Conference on Artificial Neural Networks, Proceedings
A2 - Iliadis, Lazaros
A2 - Papaleonidas, Antonios
A2 - Angelov, Plamen
A2 - Jayne, Chrisina
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 26 September 2023 through 29 September 2023
ER -