Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (19): 209-220.DOI: 10.3778/j.issn.1002-8331.2306-0407

• Graphics and Image Processing • Previous Articles     Next Articles

Small Object Detection Based on Feature Space and Coordinate Convolution

CHENG Mengyang, GE Haibo, HE Wenhao, MA Sai, AN Yu   

  1. School of Electronic Engineering, Xi’an University of Posts and Telecommunications, Xi’an 710121, China
  • Online:2024-10-01 Published:2024-09-30

基于特征空间与坐标卷积的小目标检测算法

程梦洋,葛海波,何文昊,马赛,安玉   

  1. 西安邮电大学  电子工程学院,西安  710121

Abstract: In order to solve the problem that the small object detdction is difficult to identify and the few extractable features faced by UAV high-altitude shooting, a small target detection algorithm combining feature space and coordinate convolution is proposed. The feature space module (FSM) is added to the YOLOv5 network architecture, and the convolution is used to assign adaptive weights to the receptive fields of different features, so as to enhance the feature extraction ability of the backbone network. It embeds the coordinate convolution model (CCM) in the neck of the model, accurately locates the location of the target and improves the detection accuracy of small targets in dense scenes. It deletes the original layer and adds a small target detection layer to reduce semantic loss, fully extracts the information in the shallow feature maps, and strengthens tiny objects in the high-altitude images detection performance. The experimental results show that on the VisDrone2019 dataset, the accuracy of the improved model is 4.1 percentage points higher than that of YOLOv5, and mAP@0.5 and mAP@0.5:0.95 are 4.6 percentage points and 3.2 percentage points higher respectively, thus verifying that the model proposed in this paper is effective in drone detection effective in the target scenario.

Key words: object detection, YOLOv5, small target, feature enhancement, coordinate convolution

摘要: 为解决无人机高空拍摄面临的小目标聚集不易识别、可提取特征少的问题,提出一种特征空间与坐标卷积结合的小目标检测算法。在YOLOv5网络架构中加入特征空间模块(feature spatial model,FSM),利用卷积为不同特征的感受野分配自适应权重,增强主干网络特征提取能力;将坐标卷积模块(coordinate convolution model,CCM)嵌入模型颈部,精准定位目标所在位置,提高密集场景下小目标检测精度;删减原始层并添加小目标检测层,减少语义损失,充分提取浅层特征图中信息,强化高空图像中微小目标检测性能。实验结果表明,在VisDrone2019数据集上,改进后模型的精确率较YOLOv5提高4.1个百分点,mAP@0.5和mAP@0.5:0.95分别提高4.6个百分点和3.2个百分点,从而验证了提出模型在无人机检测小目标场景中具备有效性。

关键词: 目标检测, YOLOv5, 小目标, 特征增强, 坐标卷积