BEVANet：用于实时语义分割的双边高效视觉注意力网络
摘要
实时语义分割面临两个挑战：既要设计高效的架构以捕获大感受野进行语义理解，又要精细地细化边界。视觉Transformer虽能有效建模长程依赖，但计算成本高。为此，我们提出了大内核注意力（LARGE KERNEL ATTENTION, LKA）机制。我们提出的双边高效视觉注意力网络（Bilateral Efficient Visual Attention Network, BEVANet）通过稀疏分解大可分离内核注意力（Sparse Decomposed Large Separable Kernel Attentions, SDLSKA）捕获上下文信息并提取视觉与结构特征。综合内核选择（Comprehensive Kernel Selection, CKS）机制动态适应感受野，从而进一步提升性能。深度大内核金字塔池化模块（Deep Large Kernel Pyramid Pooling Module, DLKPPM）通过融合膨胀卷积和大内核注意力丰富上下文特征。双边架构促进分支间频繁通信，边界引导自适应融合（Boundary Guided Adaptive Fusion, BGAF）模块通过在边界引导下整合空间与语义特征，增强边界描绘。BEVANet在无预训练情况下实现33 FPS的实时分割，mIoU达79.3%，在ImageNet预训练后在Cityscapes上达81.0% mIoU，展现出领先的性能。代码与模型已公开：https://github.com/maomao0819/BEVANet。
引用
@article{arxiv.2508.07300,
title = {BEVANet: Bilateral Efficient Visual Attention Network for Real-Time Semantic Segmentation},
author = {Ping-Mao Huang and I-Tien Chao and Ping-Chia Huang and Jia-Wei Liao and Yung-Yu Chuang},
journal= {arXiv preprint arXiv:2508.07300},
year = {2025}
}
备注
Copyright 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works