Feature-Level and Object-Level Scale Enhancement for Few-Shot Defect Detection

In industrial defect detection, challenges such as the scarcity of defect samples and significant variations in target scales make it difficult for conventional deep learning methods to achieve effective defect detection with limited labeled samples. To address these challenges, we propose a scale-enhanced few-shot defect detection model based on fine-tuning method. The model is first pre-trained on the Pascal VOC dataset and then fine-tuned on a defect dataset. In the feature fusion phase, our model incorporates a Bidirectional Feature Pyramid Network (BiFPN) for bidirectional multi-scale feature fusion, promoting sufficient interaction and information transfer across multiple scales. Furthermore, we design an auxiliary branch for multi-scale object enhancement to increase the diversity of object scales, thereby improving the model's ability to perceive defect object at various scales. Experiments on the NEU-DET and PKU-Market-Phone dataset demonstrate the effectiveness of the proposed approach. The results indicate that our model surpasses ten state-of-the-art few-shot object detection models, achieving a 3.6% and 1.5% improvement in mean Average Precision (mAP) on the two datasets under the 10-shot settings.