面向小目标检测的改进YOLOv3算法

An Improved YOLOv3 Algorithm for Small Target Detection

搭载在无人机平台上的目标检测系统,在实际应用中往往面临许多小目标检测任务。为了克服小目标检出率低、检测精度差的问题,提出了一种基于YOLOv3的小目标检测改进算法。首先,通过K-means聚类算法对高空视角的遥感小目标数据集进行聚类分析,重新设置锚框的个数与相应参数。然后,在特征提取网络部分重新配置5次下采样后的残差块数量,并在更浅层的网络引出一个输出尺度,将其与上一个输出尺度中的特征信息进行特征拼接,使其保留更多小目标信息和边缘信息。通过对测试数据集进行测试分析,改进后检测算法的mAP达到92.21%,相较于原YOLOv3提升了5.84%,有效解决了YOLOv3部分小目标容易漏检的问题。

The target detection system carried on UAV platform often faces many small target detection tasks in practical application.In order to overcome the problems of low detection rate and poor detection accuracy an improved small target detection algorithm based on YOLOv3 is proposed.Firstly the K-means clustering algorithm is used to conduct clustering analysis on the remote sensing small target data set from high-altitude perspective and the number of anchor boxes and corresponding parameters are reset.Then in the part of feature extraction network the number of residual blocks after five times of down sampling is reconfigured and an output scale is introduced in the shallower network whose feature information is spliced with that in the previous output scale so as to retain more small target information and edge information.Through the test and analysis of the test data set the mAP of the improved detection algorithm reaches 92.21%which is 5.84%higher than that of the original YOLOv3.It effectively solves the problem of YOLOv3 that miss detection is likely to occur on some small targets.