Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (6): 241-250.DOI: 10.3778/j.issn.1002-8331.2210-0118

• Graphics and Image Processing • Previous Articles     Next Articles

Improved Fisheye Image Target Detection Algorithm Based on YOLOv5 Network

LYU Xiaoling, YANG Shengyue, ZHANG Minglu, LIANG Ming, WANG Junchao   

  1. 1.School of Mechanical Engineering, Hebei University of Technology, Tianjin 300400, China
    2.School of Artificial Intelligence, Hebei University of Technology, Tianjin 300400, China
  • Online:2023-03-15 Published:2023-03-15



  1. 1.河北工业大学 机械工程学院,天津 300400
    2.河北工业大学 人工智能与数据科学学院,天津 300400

Abstract: The existing target detection algorithms have the problems of low detection accuracy and poor real-time performance on the fisheye image dataset. Therefore, this paper proposes a target detection algorithm in fisheye images using improved network YOLOv5 in chemical scenes. Firstly, aiming at the problem that there is no public chemical industry fisheye image database at present, the dataset is transformed into the image with the same distortion effect of the fisheye image by using the coordinate relationship of pixels between different types of images. Then, inserting line scan algorithm in YOLOv5s data preprocessing stage can effectively eliminate the redundant information of fisheye image. Finally, the attention mechanism scSE and dilated convolution are used to improve the lightweight network ShuffleNetV2, and the improved lightweight network is used to replace the main feature extraction network in the original YOLOv5s to reduce the size of the model while ensuring algorithm detection accuracy. The experimental results show that under the same conditions of experimental settings, the improved algorithm reduces the model size from 27.4 MB  to 14.2 MB, and the detection accuracy reaches 98.46% from 97.86%.

Key words: deep learning, target detection, fisheye image, ShuffleNetV2, scSE, YOLOv5

摘要: 针对目标检测算法应用在鱼眼图像数据集上检测精准率低、算法实时性差等问题,提出了在化工场景下利用改进网络YOLOv5进行鱼眼图像中的目标检测算法。由于无公开化工场景鱼眼图像数据集,提出了利用不同类型图像间像素点的坐标关系,将数据集转换为同鱼眼图像具有相同畸变效果的图像。为消除鱼眼图像中有效区域外的冗余信息,将线扫描算法应用到YOLOv5s数据预处理阶段。为在缩减模型的同时保证算法的检测精准率,提出了采用注意力机制scSE和空洞卷积来改进轻量级网络ShuffleNetV2,并利用改进后的轻量级网络代替原YOLOv5s中主特征提取网络。实验结果表明,在实验设置相同的条件下,改进后的算法在模型从27.4?MB缩减到14.2 MB的情况下,检测精准率从97.86%提高到98.46%。

关键词: 深度学习, 目标检测, 鱼眼图像, ShuffleNetV2, scSE, YOLOv5