遥感影像目标检测的轻量化密集连接网络北大核心CSCD

针对无人机遥感影像中可能出现的多个密集的小目标,在检测时会出现误检、漏检的难点问题,本文提出了一种基于YOLO v4的具有密集连接网络模块的遥感影像轻量化目标检测算法,实现了对无人机遥感影像中车辆小目标的高精度识别。首先,对YOLO v4主干网络CSP Darknet53的卷积层采用密集连接、稀疏连接两种处理方式,加强特征的提取和重复使用,以缓解梯度消失问题;然后,对此模块进行模型剪裁,减少网络层数并定义为新的密集连接网络模块;最后,在NWPU-VHR-10数据集和笔者所在课题组制作的Vehicle-850无人机影像数据集上进行了对比试验并取得了较好的效果。本文改进后的网络结构在提高遥感影像目标检测准确率的同时,缩短了网络模型的收敛时间,减少了模型占用的内存空间,提高了遥感影像中目标检测的速度。

A lightweight dense connection network for object detection of remote sensing images

Two issues arising out of the small objects, detected within UAV remote sensing images, sometimes happening to be multiple and dense increases in both the misdetection and miss out rates. A lightweight target detection algorithm is proposed based on YOLO v4 with a dense connection network module to realize high precision identification of the small targets. Firstly, feature reuse and feature extraction enhancement are facilitated by adjusting the convolutional layers of CSP Darknet53 (backbone network of YOLO v4) to work in two modes of actions:dense connection and sparse connection; the gradient disappearance problem is also alleviated. Secondly, the proposed module tailors the models to fit into a reduced number of network layers; then the resulting module is redefined as the new densely connected network module. Experiments are carried out on the NWPU-VHR-10 and Vihicle-850 UAV image datasets, a courtesy of our research group. Our model has the upper hand in terms of the number of the accurately detected small targets within the remote sensing images; our implementation is also cost-effective in terms of the network model convergence time and memory consumption. A significant improvement in the speed of detection is gained.