Abstract: Abstract: To address the issues of insufficient representation ability of category prototypes and low utilization rate of prior masks in Few-Shot Semantic Segmentation, a few-shot semantic segmentation network with prototype enhancement and prior guidance, named PEPGNet, was proposed. First, a Focused Attention Prototype Module was designed. By introducing focal modulation and self-attention mechanisms, multi-scale context and global context information were fused to dynamically weight key regions of support set features and suppress background noise, thereby generating enhanced prototypes with both global semantic consistency and local discriminability. Then, a Prior Mask Guidance Module was constructed. The limitation of using prior masks merely as positional indices was broken. By utilizing residual networks and self-attention mechanisms to deeply encode the masks, local edge details and global geometric correlations were collaboratively captured. Consequently, rough probability distributions were reconstructed into high-dimensional structure-aware features rich in spatial semantics to provide precise geometric guidance for the decoding stage. Experimental results show that in 1-shot and 5-shot FSS tasks on PASCAL and COCO datasets, compared with the baseline method ProtoFormer, the mean Intersection over Union (mIoU) of PEPGNet on the PASCAL dataset increased by 1.75 and 0.93 percentage points, respectively, and increased by 1.83 and 1.25 percentage points on the COCO dataset, respectively. Finally, visualization results indicate that by optimizing prototype construction logic and deeply mining prior knowledge, the prototype discriminative ability and segmentation accuracy of the network under few-shot conditions are significantly enhanced by PEPGNet, providing better technical support for precise segmentation in few-shot scenarios.

Key words: Keywords: few-shot semantic segmentation, prototype learning, attention mechanism, prior mask, multi-scale context, feature encoding

摘要： 摘 要: 针对小样本语义分割任务中类别原型表征能力不足以及先验掩码利用率低的问题，提出一种融合原型增强与先验引导的小样本语义分割网络（PEPGNet）。首先设计聚焦注意力原型模块，通过引入焦点调制与自注意力机制，融合多尺度上下文与全局上下文信息，动态加权支持集特征的关键区域以抑制背景噪声，从而生成兼具全局语义一致性与局部判别性的增强原型。其次，构建先验掩码引导模块，打破掩码仅作为位置索引的局限，利用残差网络与自注意力机制对其进行深度编码，协同捕捉局部边缘细节与全局几何关联，将粗糙的概率分布重构为富含空间语义的高维结构感知特征，为解码阶段提供精准的几何引导。实验结果表明，在PASCAL和COCO数据集的1-shot和5-shot的小样本语义分割任务中，相较于基线方法ProtoFormer，PEPGNet在PASCAL数据集上平均交并比（mIoU）分别提升了1.75和0.93个百分点，而在COCO数据集上分别提升了1.83和1.25个百分点。此外可视化结果表明，PEPGNet通过优化原型构建逻辑与深度挖掘先验知识，能够显著增强网络在小样本条件下的原型判别能力与分割精度，为小样本条件下的精确分割提供了更好的技术支持。

关键词: 关键词: 小样本语义分割, 原型学习, 注意力机制, 先验掩码, 多尺度上下文, 特征编码