基于改进YOLO_v3-SPP的无人机图像目标检测系统与实现

Target Detection System and Realization of UAV Image Based on Improved YOLO_v3-SPP

针对现有无人机图像目标检测算法检测率低、误检率高的问题,提出了一种改进YOLO_v3-SPP网络的目标检测算法。以YOLO_v3-SPP网络为基础,在网络结构上进行更多尺度的特征融合,将DarkNet-53中的第3、第4卷积层的信息经下采样后送入网络中进行融合,可提高目标多尺度检测的准确度;引入异常检测网络对目标进行二次判定,可剔除误检样本,有效降低目标检测的误判率。在公开数据集和自建军事目标数据集上进行实验,平均精度提升了4%,改善了现有算法在侦察影像上应用时存在的漏检和误检问题。最后将算法模型移植到硬件平台,系统检测速度小于40 ms,具有较好的检测效果。

Aiming at the problems of low detection and high false detection of existing UAV image target detection algorithms,the method of intelligent target detection is carried out.Based on YOLO_v3-SPP network,more scales of feature fusion are performed on the network structure,and the information of the third and fourth convolution layers in DarkNet-53 is down sampled and sent to the network for fusion,which can improve the accuracy of multi-scale target detection;anomaly detection network is introduced to make a second judgment on the target,which can eliminate the misdetected samples and effectively reduce the misjudgment rate of target detection.Experiments on public data sets and self-built military target data sets show that the map value is improved by 4%,which shows that the algorithm improves the problems of missed detection and misdetection in the application of existing algorithms to reconnaissance image.Finally,the algorithm model is transplanted to the hardware platform,and the system detection speed is less than 40 ms,which has a good detection effect.