基于YOLOv4的目标检测优化方法

A Lightweight Target Detection Algorithm Based on YOLOv4-GC

针对YOLOv4目标检测网络结构复杂,参数量以及计算量大等问题,提出了一种轻量化目标检测算法(YOLOv4-GC)。首先,使用ghostnet结构替换原始YOLOv4的主干网络,降低了获取冗余特征图像的计算量,在SPP与PANet模块中使用深度可分离卷积,使模型的计算量和参数量比原始YOLOv4分别降低82%和80%;再结合PyConv多尺度卷积设计出Py-PANet金字塔结构,提高了模型对于图像特征的提取和融合能力。在Pascal VOC数据集上的实验结果表明,在保证模型精度的情况下模型的参数量和计算量相比原始有明显降低。

Aiming at the problems of the complex structure of the YOLOv4 target detection network,the amount of parameters and the large amount of calculation,a lightweight target detection algorithm(YOLOv4-GC)is proposed.First,the ghostnet structure is used to replace the original YOLOv4 backbone network,and the amount of calculation for acquiring redundant feature images is reduced.The depth separable convolution is used in SPP and PANet modules to reduce the calculation and parameters of the model by 82%and 80%respectively compared with the original YOLOv4.Then combined with PyConv multi-scale convolution,Py-PANet pyramid structure is designed to improve the model’s ability to extract and fuse image features.The experimental results on the Pascal VOC data set show that the amount of parameters and calculations of the model are reduced significantly under the condition that the precision of the model is assured.