基于注意力机制改进的轻量级目标检测算法

Attention Mechanism-Based Improved Lightweight Target Detection Algorithm

针对通用的目标检测算法在检测生活场景下的多类目标时检测精度低、速度较慢的问题,提出了一种基于注意力机制改进的轻量级目标检测算法YOLOv4s。该算法以CSPDarknet53-s作为主干特征提取网络提取图像特征,通过注意力模块进行特征选择,再利用特征金字塔网络对特征进行融合,最后通过检测头分别处理特征融合后的两个输出,进而提高对生活场景下多类目标检测的能力。实验结果表明:相比改进前的算法,YOLOv4s算法在PASCAL VOC数据集上的平均均值精度(mAP)及MS COCO数据集上的平均精度(AP)都有一定程度的提升;相较于轻量级算法Efficientdet,YOLOv4s算法在MS COCO数据集上的AP也有一定提高,并且实现了有效的显著目标检测。

A lightweight YOLOv4s based on an attention mechanism is proposed to address the low accuracy and slow speed issue of the general target detection algorithm in multitarget life scenes.First,CSPDarknet53-s was used as the backbone network to extract image features,and these features were selected using the attention block.Subsequently,the feature pyramid network was adopted to fuse the features.Finally,the YOLOv4s head was used to process the two outputs after the feature fusion to improve the multitarget detection ability in living scenes.According to the experiment results,the YOLOv4s algorithm outperforms the prior algorithm in the PASCAL VOC and MS COCO datasets,exhibiting improvement in the mean average precision and average precision.Compared with the lightweight algorithm Efficientdet,the YOLOv4s algorithm also has a certain improvement in the AP on the MS COCO dataset,and achieves effective significant target detection.