基于全局语义学习和显著目标感知的激光干扰图像修复  

Laser jamming image inpainting based on global semantic learning and salient target awareness

作　　者：开志强 苗锡奎 马天磊 冯斌 艾彬 Kai Zhiqiang;Miao Xikui;Ma Tianlei;Feng Bin;Ai Bin(School of Electrical and Information Engineering,Zhengzhou University,Zhengzhou 450001,China;State Key Laboratory of Intelligent Agricultural Power Equipment,Luoyang 471032,China;Unit 63891 of PLA,Luoyang 471000,China;School of Automation,Northwestern Polytechnical University,Xi′an 710072,China;Zhengzhou Think Freely Hi-Tech Co.,Ltd,Zhengzhou 450066,China)

机构地区：[1]郑州大学电气与信息工程学院,郑州450001 [2]智能农业动力装备全国重点实验室,洛阳471032 [3]中国人民解放军63891部队,洛阳471000 [4]西北工业大学自动化学院,西安710072 [5]郑州畅想高科股份有限公司,郑州450066

出　　处：《仪器仪表学报》2024年第7期38-51,共14页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(62373330);河南省重点研发专项(241111110500);中原科技创新青年拔尖人才项目;河南省本科高校青年骨干教师培养计划(2023GGJS005)项目资助。

摘　　要：光电成像侦察装备在受到激光干扰时,成像中会出现干扰光斑。激光干扰光斑会显著降低图像质量并遮挡目标关键信息,严重影响检测与跟踪系统的性能。针对典型目标场景下的激光干扰图像,构建了一种基于全局语义学习和显著目标感知的修复网络,旨在推理出语义合理和目标完整的图像内容。提出了一种门控语义学习机制,首先通过上下文注意力机制建立干扰区域和已知区域之间的远距离信息相关性并推理干扰区域内容;然后利用多尺度特征聚合模块在不同感受野上细化推理区域的内容,实现在干扰区域重建丰富的语义信息;最后通过门控机制自适应融合已知区域和重建区域特征,提高修复图像的全局语义一致性。同时,设计了显著目标一致性损失,利用基于显著目标掩码的梯度惩罚方法,从形状和纹理两个方面指导修复网络感知显著目标,提高修复目标的轮廓清晰度和纹理连贯性。在飞机、桥梁、道路等典型目标场景下的实验结果表明,提出的网络在生成视觉真实且目标完整的内容方面优于其他方法,并在面对复杂干扰光斑时,具有很好的泛化性能。In the context of laser interference in electro-optical imaging reconnaissance equipment,interference spots often appear in the imagery.These laser jamming spots significantly degrade image quality and obscure target information,severely impacting detection and tracking systems′performance.For addressing laser jamming images in typical target scenarios,an inpainting network is developed based on global semantic learning and salient object awareness.A gated semantic learning mechanism is specifically proposed.Initially,a contextual attention mechanism is employed to establish long-range correlations between the interfered and known regions,enabling the inference of content in the interfered regions.Then,a multi-scale feature aggregation module refines the inferred content across different receptive fields,reconstructing rich semantic information in the interfered areas.Finally,a gating mechanism adaptively fuses features from the known and reconstructed regions,enhancing the global semantic consistency of the restored image.Additionally,a salient target consistency loss is designed to guide the inpainting network in perceiving salient targets,improving the sharpness of object contours and texture coherence using a gradient penalty method based on the salient target mask.Experimental results in typical target scenarios such as aircraft,bridges,and roads demonstrate that the proposed network outperforms other methods in generating visually realistic and complete content,with good generalization performance in dealing with complex interference spots.

关 键 词：图像修复 激光干扰 生成对抗网络 注意力机制 显著目标 

分 类 号：TH39[机械工程—机械制造及自动化]