摘要
为了解决目标检测模型参数冗余大、终端部署检测实时性差等问题,文中引入一种基于网络通道剪枝的模型压缩算法。针对轻量级YOLOv4-MobileNet检测模型存在的通道冗余问题,提出了被动式剪枝和主动式剪枝两种优化方案。剪枝后模型Pruned_Model参数量仅5.9M,推理速度达到324.8FPS,较原始模型压缩过5倍,加速近3倍,总体精度损失仅2.1%。结果表明,剪枝方案在道路损坏检测任务上以极小的精度损失换取了大规模的模型压缩。
In order to solve the problems of large redundancy of object detection model parameters and poor real-time performance in terminal deployment,the paper introduces a model compression algorithm based on network channel pruning.Based on the channel redundancy problem in the lightweight YOLOv4-MobileNet detection model,two optimization method,including passive pruning and active pruning are proposed.After pruning,model Pruned_Model parameter amount is only 5.9M,and the inference speed reaches 324.8FPS,which is 5 times compressed compared with the original model,and the speed is nearly 3 times faster.The overall accuracy loss is only 2.1%.The results show that the pruning scheme trades a very small loss of accuracy for large-scale model compression in the task of road damage detection.
作者
范良辰
夏颖慧
李雨诗
陈绪君
FAN Liang-chen;XIA Ying-hui;LI Yu-shi;CHEN Xu-jun(College of Physical Science and Technology,Central China Normal University,Wuhan 430079,China)
出处
《信息技术》
2022年第3期96-102,108,共8页
Information Technology