光学学报, 2019, 39 (7): 0715004, 网络出版: 2019-07-16
改进的YOLO V3算法及其在小目标检测中的应用 下载: 5901次
Improved YOLO V3 Algorithm and Its Application in Small Target Detection
机器视觉 小目标检测 YOLO V3 VEDAI数据集 K-means聚类算法 machine vision small target detection YOLO V3 VEDAI dataset K-means clustering algorithm
摘要
针对图像中小目标检测率低、虚警率高等问题,提出了一种YOLO V3的改进方法,并将其应用于小目标的检测。由于小目标所占的像素少、特征不明显,提出对原网络输出的8倍降采样特征图进行2倍上采样,将2倍上采样特征图与第2个残差块输出的特征图进行拼接,建立输出为4倍降采样的特征融合目标检测层。为了获取更多的小目标特征信息,在YOLO V3网络结构Darknet53的第2个残差块中增加2个残差单元。利用K-means聚类算法对目标候选框的个数和宽高比维度进行聚类分析。用改进的YOLO V3算法和原YOLO V3算法在VEDAI数据集上进行对比实验,结果表明改进后的YOLO V3算法能有效检测小目标,对小目标的召回率和检测的平均准确率均值都有明显的提升。
Abstract
This study proposes an improved detection algorithm of YOLO V3 specially applied in small target detection to solve the problems of low detection and high false alarm rates of small targets in an image. The resolution of small targets is low, and their features are not obvious; thus, this study proposes 2× upsampling for the feature map down-sampled by 8× of the previous network,and the feature map upsampled by 2× is concatenated with the output of the second ResNet block unit. A feature fusion target detection layer, whose feature map is down-sampled by 4×, is established. Two ResNet units in the second ResNet block unit of Darknet53 in the YOLO V3 network structure are added to obtain more features of the small target. The K-means clustering algorithm is used to select the number of candidate anchor boxes and aspect ratio dimensions. A comparative experiment is performed based on the improved YOLO V3 algorithm on the VEDAI dataset and YOLO V3 algorithm. The results show that the improved YOLO V3 algorithm can efficiently detect small targets and improve the mean average precision and recall rate of small targets.
鞠默然, 罗海波, 王仲博, 何淼, 常铮, 惠斌. 改进的YOLO V3算法及其在小目标检测中的应用[J]. 光学学报, 2019, 39(7): 0715004. Moran Ju, Haibo Luo, Zhongbo Wang, Miao He, Zheng Chang, Bin Hui. Improved YOLO V3 Algorithm and Its Application in Small Target Detection[J]. Acta Optica Sinica, 2019, 39(7): 0715004.