ES-YOLO：基于细节特征增强与冗余特征抑制的小目标检测方法

打开文本图片集

中图分类号：TP311 文献标志码：A

本文引用格式：怡，等.ES-YOLO：基于细节特征增强与冗余特征抑制的小目标检测方法[J].华东交通大学学报，2025，42（6）：42-50.

ES-YOLO： Small Object Detection Method Based on Detail Feature Enhancement and Redundant Feature Suppression

Zhu Zhiliang'， Huang Xinrong'，Liu Yil，Luo Wenjun²， Zhu Bitang²， Zhang Xiaogang² （1.SchoolofInformationand Software Engineering，East China Jiaotong University，Nanchang 33ool3，China; 2.School of Civil Engineering and Architecture，East ChinaJiaotong University，Nanchang 33oo13，China; 3.Electric Department of China Railway Wuhan Bureau Group Co.，LTD.，Wuhan 43o071，China）

Abstract：To addressthe problem ofdetailfeature loss oflow-altitude smallobjects during multi-layer down-sampling，a smallobject detection model ES-YOLO is proposed，based on detail feature enhancement and redundant feature suppresson.The method is built upon the lightweight YOLOv5s framework and constructs a dual-feature optimization mechanism consisting of spatial detail enhancement （SDE）and redundant feature suppression （RFS） modules. SDE collaborates dynamic upsampling with transposed convolution upsampling to achieve scale-adaptive finerecovery of spatial details and structural consistencyreconstruction， enhancing small object texture and boundary information.RFS models feature dependencies across both channel and spatial dimensions to suppress background noise and redundant responses， improving feature purity and object saliency.Experimental results show that ES-YOLO achieves improvements of 12.97 percentage point and 9.22 percentage point in mAP @0.5 and mAP@[0.5;0.95] ，respectively， compared to YOLOv5s on the VisDrone2019 dataset.The proposed model requires only 38.59% of the GFLOPs of YOLOv8m， achieving a significant reduction in computational cost.

Key words： small object detection; detail feature enhancement; redundant feature suppression; YOLO

Citation format：ZHU ZL ，HUANG XR ，LIUY， etal.ES-YOLO：Small objectdetectionmethod based on detail feature enhancement and redundant feature suppression[J]. Journal of East China Jiaotong University， 2025，42（6）：42-50.

在计算机视觉领域，目标检测一直是基础且关键的研究方向，广泛应用于自动驾驶、路面病害监控和遥感图像处理[3等实际场景。（剩余14647字）