<p>Recently, deep learning methods have been widely used for detecting defects in industrial environments due to their strong feature extraction and representation capabilities. Nevertheless, the effectiveness of these methods heavily relies on large amounts of labeled data, which limits their applicability in practical industrial scenarios where defect samples are scarce. Furthermore, variations in defect size and inconsistencies in the learned feature space pose additional challenges in few-shot settings. To address these issues, we propose a novel few-shot learning defect detection (FSDD) framework. Transfer learning is employed to incorporate prior knowledge from large-scale datasets, effectively reducing the demand for extensive annotations. In addition, a Spatial Adaptive Feature Pyramid Network (SAFPN) is introduced to dynamically fuse multi-scale spatial features through adaptive weighting mechanisms, alleviating performance degradation caused by defect scale variations in few-shot scenarios. To further address distribution shifts and feature inconsistencies introduced by the Distribution-Distilled Synthetic Features (DDSFs) module, a Contrastive Learning Guided Memory Bank (CGMB) is designed to store and compare feature distributions, enhancing inter-class separability and intra-class compactness. To meet the strict real-time requirements of large-scale industrial defect inspection, FSDD leverages GPU-based parallel computation to support fast and accurate inference in high-performance computing environments. Experimental results on the public NEU-DET and GC10-DET datasets demonstrate that the proposed method significantly outperforms existing approaches under various few-shot scenarios, validating its effectiveness and practical value.</p>