Few-shot semantic segmentation in complex industrial components

In the semantic segmentation of complex industrial parts, it is usually impossible to obtain enough defective industrial parts for training. This paper proposes a few sample semantic segmentation model based on complex industrial components to address this issue. This model adopts a U-shaped structure and skip connections to better learn the detailed features of high-resolution data. To minimize noise interference in this process, we introduce a gated attention module to reduce noise in the fusion of two types of feature information. Meanwhile, we propose an adaptive multi-scale attention module to better extract global information from advanced feature information. Compared with similar methods, the proposed model achieved leading results in small sample industrial datasets and demonstrated its generalization ability through the use of VOC datasets. Specifically, we achieved a performance improvement of 2.5