基于多尺度先验融合的小样本语义分割方法

Few-Shot Semantic Segmentation Method Based on the Fusion of Multi-Scale Priors

摘要: 小样本语义分割侧重于模型泛化能力的提升，用有限的样本为未见类提供像素级分割。现有方法在分割性能上已经取得了巨大的进步，但仍受到部分因素的限制，例如当同类物体中因较大的角度、尺寸等差异，存在类内差异时，将导致模型难以捕捉类别语义。因此，本文提出基于多尺度先验信息融合的小样本语义分割模块，通过提取高级特征中的多尺度先验掩码，为模型补充类别相关信息，降低类内差异的影响。为了融合多尺度特征及其先验信息，本文提出权重自适应特征融合模块，为各尺度添加信息交互，并加权组合特征。最终通过消融实验以及与基类方法和其他经典算法的对比实验，证明了本方法的有效性。

Abstract: Few-shot semantic segmentation focuses on improving the model’s generalization ability to provide pixel-level segmentation for unseen classes with limited samples. While existing methods have made significant progress in segmentation performance, they are still constrained by certain fac-tors. For instance, when there are intra-class gaps due to significant variations in angles, sizes, etc., within the same class, the model struggles to capture category semantics. Therefore, this paper proposes a few-shot semantic segmentation module based on the fusion of multi-scale prior infor-mation. By extracting multi-scale prior masks from high-level features, this module supplements category-related information for the model, reducing the impact of intra-class differences. To inte-grate multi-scale features and their prior information, the paper introduces a Weighted Adaptive Feature Fusion module, adding information interaction across various scales and combining features with weighted contributions. Ultimately, through ablation experiments and comparative ex-periments with baseline methods and other classical algorithms, the effectiveness of this approach has been demonstrated.