Abstract: To address the issue of insufficient accuracy and robustness in existing feature extraction methods due to the lack of large datasets in the intelligent design of power equipment, a WinCLIP+ method based on small-sample learning is proposed to enhance feature extraction performance under zero-sample and small-sample conditions. This method combines the advantages of both zero-sample and small-sample learning,utilizes a pre-trained CLIP model for multi-scale feature extraction,introducesa small number of design reference samples, and incorporatea reference correlation module alongside a multi-scale feature fusion mechanismto enhance the extraction capability of features of different types and scales. Furthermore, WinCLIP+ further enhances the robustness of feature extraction by integrating linguistically guided predictions and visual reference information. To verify the effectiveness of this approach,an anomaly detection task is conducted. The experimental results indicate that WinCLIP+ achieves a detection accuracy of 84% under zero-sample conditions and 91% under small-sample conditions, demonstrating its significant effectiveness and robustness for feature extraction in the intelligent design of power equipment.

Key words: small-sample learning, anomaly detection, image segmentation, power equipment defects, feature extraction