摘要
为高效应用各类模态不同视角的图像信息,降低精导武器的保障要求,充分发挥其实战效能,需进一步提升景象匹配方法的适应性及抗干扰能力。聚焦多模态多视角景象匹配技术,重点探索可解释性强的基于深度学习的景象匹配方法。该方法整合自动编解码网络和孪生网络,通过区分场景图像的高低频信息及改进相似性测度对比方式,提升景象匹配的多模态适应性及多场景泛化能力;在构建景象匹配方法的基础上,进一步分析其在精确制导武器作战中的应用方式,展望兼容语言模态与图像信号的景象匹配新方向。突破高可靠高适应性的多模态多视角景象匹配技术能够增强精导武器的抗扰能力和复杂场景适应性,更进一步释放精导武器作战效能。
In current complex battlefield environments,where adversarial interference is becoming increasingly intense,navigation and guidance methods based on a single imaging modality struggle to fully exploit the practical effectiveness of precision-guided weapons.To efficiently utilize image information from various modalities and angles,and to reduce the logistical requirements of precision-guided weapons,the adaptability and anti-interference capabilities of scene matching methods must be enhanced.The multi-modal and multi-view scene matching techniques are focused on and a deep learning-based method with strong interpretability is explored.This approach integrates autoencoder-decoder networks and Siamese networks,and it enhances the multi-modal adaptability and cross-scene generalization capability of scene matching by distinguishing the high-frequency and low-frequency information of scene images and improving the similarity measurement.Concurrently,while developing the scene matching method,this research analyzes the matching strategies in precision guidance applications and envisions novel directions for scene matching that incorporate both linguistic modalities and image information.The breakthrough in highly reliable and adaptive multi-modal multi-view scene matching can bolster the anti-interference capabilities and adaptability of precision-guided weapons in complex scenarios,thereby further unleashing the effectiveness of precision-guided weapons.
作者
滕锡超
叶熠彬
刘学聪
TENG Xichao;YE Yibin;LIU Xuecong(College of Aerospace Science and Engineering,National University of Defense Technology,Changsha 410073,China)
出处
《国防科技》
2023年第5期26-34,共9页
National Defense Technology
基金
国家自然科学基金项目(61801491)。
关键词
景象匹配
多模态多视角
特征提取
深度学习
精确制导
scene matching
multi-modal and multi-view
feature extraction
deep learning
precision guidance
