基于改进的YOLOv4-tiny模型剪枝与量化

Pruning and Quantization Based on Improved YOLOv4-tiny

针对YOLOv4-tiny存在计算量较大,检测精度低,无法满足嵌入式设备实时性需求的问题,论文基于MobileNetv3改进的轻量级网络YOLOv4-E,使用BN层的γ尺度因子对冗余的特征通道进行剪枝,在25%剪枝率下模型大小降低到了6.7MB,mAP仅降低了0.59%,FPS提升了8.8%。同时使用NCNN前向推理框架对剪枝后的模型进行Int8量化,在树莓派4B上检测单张图片仅需173 ms,满足了实时性需求。

Aiming at the problem that YOLOv4-tiny has a large amount of calculation and low detection accuracy,which cannot meet the real-time requirements of embedded devices,based on the improved lightweight network YOLOv4-E of MobileNetv3,this paper uses the scale factor of BN layer to prune the redundant characteristic channels.At 25% pruning rate,the size of the model is reduced to 6.7 MB,mAP is reduced by only 0.59%,and FPS is increased by 8.8%.At the same time,the NCNN forward reasoning framework is used to quantify the Int8 of the pruning model.It is only 173 ms to detect a single image on RaspberryPi 4B,which meets the real-time requirements.