跨层级交互与方位感知的航拍图像语义分割

打开文本图片集

关键词：跨层级交互；语义分割；航拍图像；轻量化；方向解耦中图分类号：TP394.1;TH691.9 文献标识码：Adoi：10.37188/OPE.20263402.0267 CSTR：32169.14.OPE.20263402.0267

Semantic segmentation of aerial images based on multi-scale feature interaction and fusion

LIU Jie¹，WU Ziyu1，TIANMing²，HAN Ke³* （1. Harbin University of Science and Technology， Key Laboratory of Pattern Recognition and Information Perception in Heilongjiang Province，Harbin 15Oo8O，China; 2.Heilongjiang Branch of China Telecom Corporation Limited，Harbin 15OO4O， China; 3. Harbin University of Commerce， School of Computer and Information Engineering， Harbin150080，China） * Corresponding author，E-mail： hanke@hrbcu. edu. cn

Abstract：To address the issues of single-scale feature extraction，detail loss，and blurrd boundaries in aerial image semantic segmentation，this paper proposed an aerial image semantic segmentation network with cross-level interaction and orientation awareness.A position awareness module was constructed through a direction-decoupled atention strategy to enhance the model's ability to process spatial directional information；a cross-level interaction module was designed for inter-channel feature interaction and fusion to improve spatial perception，while a channel-spatial attention mechanism was used to enhance feature extraction capabilities and alleviate detail bluring issues in complex scenes；finally，a lightweight design was implemented for the segmentation head，removing redundant operations to reduce computational load while ensuring segmentation performance. Experimental results indicate that the proposed network achieves a 1.7% and 1.3% improvement in mean intersection over union on the UAVid and Aeroscapes datasets，respectively，compared to the baseline model SegFormer，demonstrating the network's effectiveness in semantic segmentation under complex conditions such as aerial images. The segmentation accuracy of the Human category improved by 1.8% compared to the baseline model，demonstrating that the network proposed in this paper performs excellently in smallobject segmentation. Compared with several mainstream networks，the method proposed in this paper achieves the highest segmentation accuracy on both datasets， showing superior generalization capability.

Key words： inter-level interaction；semantic segmentation；aerial imagery； lightweighting； direction decoupling

1引言

语义分割技术是一种图像处理技术，其目标是将图像中的每个像素分配到预定义的类别中，实现像素级的分类[1]。（剩余17101字）