GH-QFL: Enhancing Industrial Defect Detection Through Hard Example Mining

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).