Abstract: In intelligent industrial production, defect detection technology using deep learning faces challenges such as insufficient defect samples, different defect sizes, and low detection accuracy. To address these problems, we propose a surface defect detection model leveraging few-shot learning, building on the Faster R-CNN(Faster Region-based Convolutional Neural Network). First, we enhance the traditional convolution in the ResNet101 and FPN(Feature Pyramid Network) backbone network with deformable convolutions to extract features. Objects are then extracted from images to generate an object pyramid, selecting corresponding features as positive samples to enrich the scale space of small samples. Finally, we encode RoI(Region of Interest) features using contrastive learning to measure the similarity between regional proposals, achieving a more compact feature representation and reducing misclassification issues in small samples. Finally, comparative experiments on the collected small sample defect dataset demonstrate the model's effectiveness, yielding a 96.6% accuracy and 70.6% average accuracy, outperforming other models.