Pruned-YOLO: Learning Efficient Object Detector Using Model Pruning

ARTIFICIAL NEURAL NETWORKS AND MACHINE LEARNING - ICANN 2021, PT IV(2021)

引用 8|浏览21
暂无评分
摘要
Accurate real-time object detection plays a key role in various practical scenarios such as automatic driving and UAV surveillance. The memory limitation and poor computing power of edge devices hinder the deployment of high performance Convolutional Neural Networks (CNNs). Iterative channel pruning is an effective method to obtain lightweight networks. However, the channel importance measurement and iterative pruning in the existing methods are suboptimal. In this paper, to measure the channel importance, we simultaneously consider the scale factor of batch normalization (BN) and the kernel weight of convolutional layers. Besides, sparsity training and fine tuning are combined to simplify the pruning pipeline. Notably, the cosine decay of sparsity coefficient and soft mask strategy are used to optimize our compact model, i.e., Pruned-YOLOv3/v5, which is constructed via pruning YOLOv3/v5. The experimental results on the MS-COCO and VisDrone datasets show that the proposed model achieves a satisfactory balance between computational efficiency and detection accuracy.
更多
查看译文
关键词
Model compression, Channel pruning, Object detection, YOLO
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要