基于深度神经网络的迷彩目标发现仿真学习方法

Simulation Learning Method for Discovery of Camouflage Targets Based on Deep Neural Networks

针对自动化迷彩目标发现学习中有效样本严重不足的问题,借鉴AlphaGo的技术思想,提出了一种基于样本模拟的深度神经网络仿真训练方法。建立了迷彩场景仿真合成模型,通过设计图像空间的复合算法、场景图像深度特征提取策略、目标融合度测量策略,以及图聚类采样算法,批量化地生成了可用于深度神经网络训练和学习的具有代表性的迷彩场景仿真样本;设计了基于深度残差神经网络的迷彩目标发现模型,并引入了多尺度网络训练方法。模拟样本和真实场景图像的实验结果表明,所提方法可有效应用于迷彩目标的自动化识别与评估。

Aiming at the problem of serious lack of effective samples in the automatic discovery of camouflage targets,a simulation training method is proposed based on the sample simulation of a deep neural network and the technical idea of AlphaGo.A simulation synthesis model of camouflage scenes is established.The compound algorithm in the image space,the deep feature extraction strategy of scene images,the measurement strategy of target fusion degree,and the sampling algorithm for graph clustering are designed,respectively.Thus the representative samples for camouflage scene simulation are batch generated,which can be used for the deep neural network training and learning.Moreover,a discovery model of camouflage targets is designed based on a deep residual neural network,in which a multi-scale network training strategy is considered.The experimental results on the simulated samples and real scene images show that the proposed method can be effectively used for the automatic discovery and evaluation of camouflage targets.