弹载融合图像深度卷积网络视觉解释  

Missile Borne Fusion Image Visual Explanations for Deep Convolutional Networks

作　　者：薛松 钱立志 杨传栋 XUE Song;QIAN Lizhi;YANG Chuandong(Department of Weapons Engineering,Army Academy of Artillery and Air Defense,Hefei 230031,China;High Overload Ammunition Guidance Control and Information Perception Laboratory,Army Academy of Artillery and Air Defense,Hefei 230031,China;Postgraduate Team,Army Academy of Artillery and Air Defense,Hefei 230031,China)

机构地区：[1]陆军炮兵防空兵学院兵器工程系,合肥230031 [2]陆军炮兵防空兵学院高过载弹药制导控制与信息感知实验室,合肥230031 [3]陆军炮兵防空兵学院研究生队,合肥230031

出　　处：《弹箭与制导学报》2022年第5期102-107,共6页Journal of Projectiles,Rockets,Missiles and Guidance

摘　　要：近年来,卷积神经网络的决策过程受到了越来越多的关注,其内部运行机制促使研究者们开展了深入研究,并形成了基于显著性映射的视觉解释理论方法。文中提出一种适用于弹载融合图像的深度卷积网络视觉解释方法,该方法通过置信度提升重组神经网络梯度映射,并结合权重参数获得显著图。实验结果表明,与经典的视觉解释方法相比,文中方法具有良好的主观视觉效果,在平均下降和平均提升两类指标上都达到了最优,同时具备较为准确的定位能力。In recent years,the decision-making process of convolutional neural networks has attracted more and more attention.Its internal operating mechanism prompts researchers to conduct in-depth research,and forms a visual interpretation theory method based on saliency map.The paper presents a deep convolutional network vision explanations method for missile borne fusion images.In this method,the gradient map of neural network is reconstructed by the“increase of confidence”,and the saliency map is obtained by combining the weight parameters.The experimental results show that compared with the classical visual interpretation methods,the paper method has good subjective visual effect,and achieves the best in the two indicators of average drop and average increase,and has more accurate positioning ability.

关 键 词：融合图像 显著性映射 卷积神经网络 计算机视觉 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]