航空兵器2025，Vol.32Issue(5)：72-82,11.DOI:10.12132/ISSN.1673-5048.2025.0063
基于KAN结构的YOLOv10s改进算法在无人机目标检测中的应用研究
Study on an Improved YOLOv10s Algorithm with KAN Structure for UAV Object Detection
摘要
Abstract
Faced with challenges such as small object sizes,weak features,heavy occlusion,large scale variation,and strong background interference under UAV perspectives,this study investigates the applicability of the Kolmogorov-Arnold Network(KAN)and its extended structures in UAV-based small object detection.Based on YOLOv10s,three improved models are proposed.Firstly,KANConv based on multi-order spline trans-formations is introduced into the Bottleneck of the C2f module,constructing the C2f_KAN module to achieve nonlinear feature mapping and detail enhancement.Secondly,Element-wise multiplication is performed on B-spline features,followed by Instance Normalization to avoid excessive activation,and a learnable parameter α is introduced to dynamically regulate the strength of spline2,forming the C2f_MultKAN module to optimize feature interaction.Thirdly,Three variants of KANConv with improved basis functions are used to replace the original convolution structures,aiming to enhance modeling capacity and reduce the computational overhead caused by B-spline operations.Experiments conducted on the VisDrone2019 dataset demonstrate that the introduction of KAN significantly enhances the original model's ability to detect small objects across various typical scenarios.Specifically,KAN-YOLO is well-suited for detecting small targets at long distances with regular spatial distribu-tion.MultKAN-YOLO performs robustly in densely populated and heavily occluded regions.JKAN-YOLO is more effective under challenging conditions such as poor illumination and strong background interference.More-over,basis function optimization reduces the computational cost by 9.8％and improves mAP@50 by 2％.This study verifies the effectiveness of the KAN architecture in complex object detection tasks.关键词
小目标检测/KAN/MultKAN/Jacobi核/YOLOv10sKey words
small object detection/KAN/MultKAN/Jacobi kernel/YOLOv10s分类
计算机与自动化引用本文复制引用
李倩倩,范军芳,徐小斌,纪毅..基于KAN结构的YOLOv10s改进算法在无人机目标检测中的应用研究[J].航空兵器,2025,32(5):72-82,11.基金项目
国家自然科学基金项目(52402442) （52402442）
十四五高水平科研团队建设支持计划(BPHR20220123) （BPHR20220123）
中国科协青年人才托举工程(YESS20230098) （YESS20230098）
北京市科协青年人才托举工程(BYESS2023310) （BYESS2023310）
