Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (20): 158-166.DOI: 10.3778/j.issn.1002-8331.2305-0004

• Graphics and Image Processing • Previous Articles     Next Articles

Algorithm of Reconstructed SPPCSPC and Optimized Downsampling for Small Object Detection

QI Xiangming, CHAI Rui, GAO Yimeng   

  1. School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2023-10-15 Published:2023-10-15

重构SPPCSPC与优化下采样的小目标检测算法

齐向明,柴蕊,高一萌   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105

Abstract: A detection algorithm is proposed of reconstructed SPPCSPC and optimized downsampling for small objects based on YOLOv7. This algorithm aims to address the challenges of detecting small objects in images, including mutual occlusion, complex backgrounds, and a limited number of feature points. To improve the detection of densely packed small objects, enhancements in the concerned dense target area are made, including cropping the CBS layer, introducing the SimAM attention mechanism, and reducing the pooling core in the SPPCSPC module of the backbone network. These modifications allow for better feature extraction of small targets that are mutually occluded. In the neck network, the SConv in the down-sampling structure is replaced by the SPD Conv and adds a quadruple down-sampling branch. These changes reduce feature loss and increase the capturing of small target features in complex backgrounds. Additionally, the Wise IoU loss function of the network model is substituted for CIoU, which focuses on the general quality frame and improves the convergence speed. Comparative and ablation experiments are conducted on the public dataset VisDrone2021, where the article increases mAP by 5.09 percentage points, achieves an FPS value of 40 and reduces the parameter count by 2.5 MB compared to the original YOLOv7 algorithm. It clearly illustrates that the modified algorithm significantly improves detection accuracy while maintaining fast inference speed and reducing the number of parameters. Furthermore, a generalization experiment is performed on the public dataset VOC2007+2012 where the mAP increased by 3.35 percentage points, indicating that the improved algorithm is versatile and can be applied to a wide range of scenarios.

Key words: small target detection, reconstructed SPPCSPC, optimized downsampling, Wise IoU, YOLOv7

摘要: 针对小目标图像检测中存在相互遮挡、背景复杂和特征点少的问题,基于YOLOv7提出一种重构SPPCSPC与优化下采样的小目标检测算法。在骨干网络的SPPCSPC模块中裁剪CBS层、引入SimAM注意力机制并缩小池化核,以提高关注密集目标区域,提取更多相互遮挡的小目标特征;在颈部网络中,将下采样结构中的SConv替换为SPD Conv,再添加一个四倍下采样分支,以减少小目标特征丢失,提高复杂背景下小目标特征捕获量;把网络模型的损失函数由CIoU替换为Wise IoU,聚焦一般质量瞄框,提升收敛速度。在公开数据集VisDrone2021上做对比实验和消融实验,该算法与原始YOLOv7算法相比,mAP提升5.09个百分点,FPS值达到40,参数量减少2.5?MB,表明小目标检测精度显著提升,同时保持了推理速度并减少了参数量;在公开数据集VOC2007+2012上做泛化实验,mAP提升3.35个百分点,表明该算法具有通用性。

关键词: 小目标检测, 重构SPPCSPC, 优化下采样, Wise IoU, YOLOv7