计算机应用

• 人工智能与仿真 •    下一篇

基于改进YOLOv4的轻量化目标检测算法

钟志峰,夏一帆,周冬平,晏阳天   

  1. 湖北大学 计算机与信息工程学院,武汉 430062
  • 收稿日期:2021-05-10 修回日期:2021-09-22 发布日期:2021-09-22 出版日期:2021-09-24
  • 通讯作者: 夏一帆

Lightweight object detection algorithm based on improved YOLOv4

  • Received:2021-05-10 Revised:2021-09-22 Online:2021-09-22 Published:2021-09-24

摘要: 针对当前YOLOv4目标检测网络结构复杂、参数多、训练所需的配置高以及实时检测每秒传输帧数(FPS)低的问题,提出一种轻量化目标检测算法(ML-YOLO)。首先,用Mobilenetv3结构替换YOLOv4的主干特征提取网络,通过Mobilenetv3中的深度可分离卷积大幅减少主干网络参数量;然后,用简化的加权双向特征金字塔(Bi-FPN)结构替换YOLOv4的特征融合网络,用Bi-FPN中的注意力机制优化目标检测精度;最后,通过YOLOv4的解码算法,生成最终的预测框,实现物品的目标检测。在VOC2007数据集的实验结果表明,ML-YOLO算法的平均准确率(mAP)达到80.22%,与YOLOv4算法相比低了3个百分点,与YOLOv5m算法相比提升了3个百分点;而模型大小仅为44.75M(Mega),与YOLOv4算法相比减少了199.5M,与YOLOv5m算法相比,只高了2.8M。所提的ML-YOLO模型,一方面较YOLOv4模型大幅减小了模型大小,另一方面,保持了较高的检测精度,表明该算法可以满足移动端或者嵌入式设备进行目标检测的轻量化和准确性需求。

关键词: 目标检测, 轻量化网络, YOLOv4, Mobilenetv3, 加权双向特征金字塔

Abstract: YOLOv4(You Only Look Once version 4) object detection network has complex structure, many parameters, high configuration required for training and low FPS(Frames Per Second) for real-time detection.In order to solve the above problems, a lightweight object detection algorithm ,ML-YOLO(Mobilenetv3Lite-You Only Look Once), based on YOLOv4 was proposed. Firstly, Mobilenetv3 was used to replace the backbone feature extraction network of YOLOv4 , which greatly reduced the amount of backbone network parameters through the deep separable convolution in Mobilenetv3; Then, a simplified Bi-FPN(weighted Bi-directional Feature Pyramid Network) structure was used to replace the feature fusion network of YOLOv4 .The object detection accuracy was optimized by the attention mechanism in Bi-FPN; Finally, the final prediction frame was generated through the YOLOv4 decoding algorithm to achieve the detection of the objects . The experimental results on the VOC(Visual Object Classes)2007 data set show that, the mAP(mean Average Precision) of the ML-YOLO algorithm reaches 80.22%, which is 3 percentage points lower than that of the YOLOv4 algorithm, and 3 percentage points higher than that of the YOLOv5m algorithm; In addition, the model size is only 44.75M ,compared with the YOLOv4 algorithm, it is reduced by 199.5M, and compared with the YOLOv5m algorithm, it is only 2.8M higher. The results prove that the proposed ML-YOLO model, on the one hand, greatly reduces the size of the model compared with the YOLOv4 model, on the other hand, it maintains a higher detection accuracy, indicating that the algorithm can meet the requirements of lightweight and accuracy for mobile or embedded devices.

Key words: object detection , lightweight network, YOLOv4( You Only Look Once version 4), Mobilenetv3, Bi-FPN( weighted Bi-directional Feature Pyramid Network)

中图分类号: