针对嵌入式设备的YOLO目标检测算法改进方法

Improvement methods for YOLO object detection algorithm targeting embedded devices

针对算法在资源有限的嵌入式设备实现困难的问题,本文基于YOLO系列算法提出适应嵌入式设备实现的轻量化改进方法。方法具体包括:基于YOLOv4-Tiny算法结构,引入GhostNet思想改进其网络主干,大量降低网络参数量和计算量;通过加强颈部网络特征融合效果,减少模型压缩导致的精度损失;采用训练中量化的方式将网络模型参数从32位浮点型数据转换为适合嵌入式设备计算的8位定点型参数。实验结果表明,改进后的网络在检测精度满足应用要求的情况下,模型尺寸相对原算法降低57%,在嵌入式设备上实现功耗仅3.795W。

To address the problem of implementing algorithms on resource-limited embedded devices,a lightweight improvement is proposed based on the YOLO series of algorithms to adapt to embedded device implementation,specifically including:improving the network backbone by introducing GhostNet ideas based on the YOLOv4-Tiny algorithm structure to significantly reduce network parameters and computational complexity;strengthening the fusion effect of neck network features to reduce accuracy loss caused by model compression;and using quantization during training to convert network model parameters from 32-bit floating-point data to 8-bit fixed-point parameters suitable for embedded device computation.Experimental results show that after the improvement in this paper,the network's model size relative to the original algorithm is reduced by 57%when the detection accuracy meets application requirements,and the power consumption for embedded device implementation is only 3.795W.