Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (17): 205-211.DOI: 10.3778/j.issn.1002-8331.2211-0298

• Graphics and Image Processing • Previous Articles     Next Articles

Small Object Detection of Improved Lightweight CenterNet

WANG Yingbo, LIU Rongxia   

  1. 1.School of Software, Liaoning Technical University, Huludao, Liaoning 125105, China
    2.School of Innovation and Practice, Liaoning Technical University, Fuxin, Liaoning 123000, China
  • Online:2023-09-01 Published:2023-09-01



  1. 1.辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105
    2.辽宁工程技术大学 创新实践学院,辽宁 阜新 123000

Abstract: Small object detection has always been a difficult point in object detection, which is characterized by small perception field and insufficient semantic features. Compared with traditional object detection algorithms, it is difficult to achieve lightweight in industrial applications. In order to improve the accuracy and speed of small object detection and reduce the amount of computation and cost, a lightweight detector SFPN-CenterNet based on CenterNet is proposed. The lightweight deep separable convolution network is used to replace the original ordinary convolution. It simplifies the FPN network, reduces the number of down-sampling layers for feature extraction and fusion, and discards high-level features that have no significant effect on small object detection. The loss function is improved to optimize the original formula and super parameters. Comparing experiments on self-made data sets, and the results show that the amount of network parameters is reduced to 1/480 by using deep separable convolution as the convolution block for feature extraction. The improved loss function reduces the false detection rate and missed detection rate of small targets. Comparing with the original algorithm, the AP of the improved algorithm increased by 3.5 percentage points and the detection speed increased by 2.74 ms.

Key words: CenterNet, small object detection, depth separable convolution, lightweight

摘要: 小目标检测一直是目标检测中的难点,其特点为感受视野小,无法获取足够的语义特征,相比传统目标检测算法在工业应用中难以实现轻量化。为提升小目标检测精度、计算速度以及减少计算量和成本开销,提出一种基于CenterNet的轻量化的检测器SFPN-CenterNet。采用轻量级的深度可分离卷积网络来替代原始的普通卷积;简化FPN网络,减少下采样层数对特征进行提取融合,舍弃对于小目标检测无显著作用的高层特征;改进损失函数,对原来的公式以及超参数进行优化。在自制的数据集上进行对比实验。结果表明:利用深度可分离卷积作为提取特征的卷积块,可以使网络参数量减少到原来的1/480;改进损失函数降低了小目标的误检率和漏检率。相对于原始算法,改进算法的AP提升了3.5个百分点,检测速度提高了2.74 ms。

关键词: CenterNet, 小目标检测, 深度可分离, 轻量化