During the manufacturing and assembly processes of carbon fiber truss platforms, various defects can arise, which may significantly impact their service life. However, the collection of defect samples is a challenging task, and the available sample size is often limited. Additionally, the types of defects are varied, and the background is complex. This paper proposes a surface defect segmentation method for carbon fiber multilayer boards based on an enhanced U-Net architecture. The VGG16 model, pre-trained on the ImageNet dataset, is incorporated, with part of its convolutional layers used as the decoder in the U-Net. The encoder incorporates a hybrid module of channel attention and spatial attention (SCSE) to effectively capture critical features while suppressing irrelevant ones, thereby constructing a novel defect segmentation network. Experimental results demonstrate that the proposed network achieves superior performance in defect detection tasks for carbon fiber materials, significantly enhancing both the accuracy of defect region identification and the detection rate of missed defects.