面向军事图像识别网络FDRCN设计及实现

Designand implementation of FDRCN for military image recognition network

随着军事目标的隐蔽性和机动性越来越好,导致检测识别难度加大,为了更加快速准确检测识别目标以防贻误战机,设计了一种新的深度网络识别方法(FDRCN)。首先通过将特征金字塔网络(FPN)和稠密性卷积神经网络(DenseNet)进行融合,构建特征金字塔稠密网络(FPDN)对目标进行高质量的特征提取;再通过RPN网络进一步确定检测目标的特征位置信息并形成目标候选区域;最后借助FCN和DenseNet的跳跃连接形成FCDN网络,实现目标种类的预测和分类并给出预测概率。结果表明,FDRCN算法模型可以大大提升检测识别性能,Box-mAP达到45.1%,Mask-mAP达到41.1%。

With the concealment and mobility of military targets are getting better and better.It makes the detection and identification more difficult,in order to detect and recognize targets more quickly and accurately,and to prevent the aircraft from being missed,a new deep network recognition method(FDRCN)isdesigned.First,construct FPDN to extract high-quality features of the targetby fusing FPN and dense net.Then use the RPN network to further determine the feature location information of the detected target and form the target candidate area.Finally,the jump connection between FCN and dense net forms the FCDN network to predict and classify the target type and give the prediction probability.The results show that the FDRCN algorithm model can greatly improve the detection and recognition performance,with Box-mAP reaching 45.1%and Mask-mAP reaching 41.1%.