Abstract: To address the challenges of complex backgrounds in small-sample defect images, making it difficult to distinguish fine-grained features, and the inability to directly apply pre-trained features for defect classification, an adaptive feature filtering and reorganization method for few-shot surface defect classification was proposed. A dual-branch parallel adapter is constructed between the pre-trained feature extraction module and the classifier to achieve precise feature enhancement: First, channel attention was used to capture cross-channel interaction information, suppressing redundant background noise while maintaining dimensionality; second, multi-path parallel convolution was employed to structurally reorganize the feature sequence, strengthening the discriminative power of local and global semantics; finally, an adaptive fusion mechanism was introduced to dynamically adjust the contribution weights of the two branches. In the 5-shot tasks on the NEU-DET, PCB, and MTD datasets, the proposed method achieves accuracies of 91.3%, 64.3%, and 80.8%, respectively, outperforming current mainstream small-sample classification models such as CLIP-ProtoNet and CLIP-Adapter. In the 1-shot and 5-shot tasks on the MVTec-FS dataset, the average accuracy of the proposed method is improved by 2.1 and 2.3 percentage points compared to the benchmark method MVREC (Multi-View Region Context), respectively. Experimental results show that the proposed method effectively improves the classification accuracy of surface defects in small samples and exhibits good generalization ability.

Key words: few-shot learning, defect classification, dual-branch structure, channel attention, feature fusion

摘要： 针对小样本缺陷图像背景复杂导致细粒度特征难区分，且预训练特征难以直接用于缺陷分类的问题，提出自适应特征过滤与重组的小样本表面缺陷分类方法，通过在预训练特征提取模块与分类器间构建双分支并行适配器，实现特征的精准增强。首先，利用通道注意力捕捉跨通道交互信息，在保持维度的同时抑制冗余背景噪声；其次，采用多路并行卷积对特征序列进行结构化重组，强化局部与全局语义的判别性；最后，引入自适应融合机制动态调节双分支贡献权重。在NEU-DET、PCB和MTD数据集的5-shot任务中，所提方法的准确率分别达到91.3%、64.3%和80.8%，优于CLIP-ProtoNet和CLIP-Adapter等当前主流小样本分类模型；在MVTec-FS数据集上的1-shot和5-shot任务中，所提方法的平均准确率相比基准方法MVREC（Multi-View Region Context）分别提高了2.1和2.3个百分点。实验结果表明，所提方法有效提高了小样本表面缺陷的分类精度，并展现出良好的泛化能力。

关键词: 小样本学习, 缺陷分类, 双分支结构, 通道注意力, 特征融合