天津科学技术出版社
多尺度特征融合在图像语义分割中的应用
陈金双 ,曹会念
作者信息
  1.  重庆工商职业学院  重庆  401520

摘要

图像语义分割作为计算机视觉的核心任务,需同时实现像素级分类精度与全局语义一致性。多尺度特征融合技术通过整合不同层级的视觉特征,有效解决了单一尺度特征难以兼顾细节信息与语义理解的瓶颈。本文系统梳理多尺度特征融合的核心原理与技术演进,分析金字塔池化、空洞卷积、注意力机制等主流融合策略的设计逻辑,探讨典型网络架构的特征融合机制,并结合实际应用场景总结技术优势。研究表明,多尺度特征融合通过互补局部细节与全局上下文信息,显著提升了复杂场景下语义分割的鲁棒性与准确性,为后续技术优化提供了理论参考与实践方向。

Multi-Scale Feature Fusion for Image Semantic Segmentation
Chen Jin-shuang, , Cao Hui-nian
Author Information
  1.  Chongqing Technology and Business Institute, Chongqing 401520, China

Abstract

Image semantic segmentation, a core task in computer vision, demands both pixel-level classification accuracy and global semantic consistency. Multi-scale feature fusion integrates visual cues from different layers, overcoming the bottleneck of single-scale representations that struggle to balance fine details with high-level semantics. This paper systematically reviews the underlying principles and technical evolution of multi-scale feature fusion, analyzes the design rationale of mainstream strategies such as pyramid pooling, atrous convolution, and attention mechanisms, examines the fusion schemes of representative network architectures, and summarizes their advantages in real-world applications. Evidence shows that complementary local details and global context provided by multi-scale fusion significantly improve the robustness and accuracy of semantic segmentation in complex scenes, offering theoretical guidance and practical directions
for future optimization.