基于改进YOLOv5s的轻量化目标检测算法

A Lightweight Object Detection Algorithm Based on Improved YOLOv5s

针对当前YOLOv5s的颈部特征提取网络PANET的特征提取不足、常规卷积Conv消耗了大量的参数量和计算量的问题,提出一种轻量化目标检测算法(RFBG-YOLO)。首先,为了提升检测器识别效果,提出多分支空洞卷积结构RFB-Bottleneck来提升PANET的特征提取能力,提高模型检测精度;然后,为了使模型更加轻量化,引入GhostConv卷积减少模型参数量,提高检测速度。在PASCAL VOC数据集上的结果表明,在检测速度影响很小的情况下,RFBG-YOLO算法的mAP@0.5为80.3%,与YOLOv5s算法相比提高了2.2个百分点,mAP@0.5∶0.95为55.1%,与YOLOv5s算法相比提高了4.2个百分点,模型参数量为5.2 MiB,与YOLOv5s算法相比降低了2.0 MiB,因此提出的RFBG-YOLO算法在保证模型轻量化的同时,具有足够高的检测精度,可以满足在轻量化目标检测场景下检测准确度的要求。

Aiming at the insufficient feature extraction of the neck feature extraction network PANET of current YOLOv5s,and the conventional convolution Conv consumes a large amount of parameters and calculations,a lightweight target detection algorithm(RFBG-YOLO) is proposed.Firstly,in order to improve the recognition effect of the detector,a multi-branch atrous convolution structure RFB-Bottleneck is proposed to improve the feature extraction ability of PANET,thereby improving the detection accuracy of the model.Then,in order to make the model more lightweight,GhostConv convolution is introduced to reduce the amount of model parameters and improve the detection speed.The results on the PASCAL VOC data set show that in the case of small impact on the detection speed,the mAP@0.5 of the RFBG-YOLO algorithm is 80.3%,an increase of 2.2 percentage points compared with that of YOLOv5s algorithm;mAP@0.5∶0.95 is 55.1%,an increase of 4.2 percentage points compared with that of YOLOv5s algorithm;the amount of model parameters is 5.2 MiB,a reduction of 2.0 MiB compared with that of YOLOv5s algorithm.Therefore,the proposed RFBG-YOLO algorithm has a high enough detection accuracy while ensuring the light weight of the model,which can meet the requirements of detection accuracy in the scenario of lightweight target detection.