基于深度学习和双域融合的红外成像制导系统复杂背景噪声去除方法  

A Denoising Method for Complex Background Noise of Infrared Imaging Guidance System Based on Deep Learning and Dual-domain Fusion

作　　者：栗苹[1,2] 周宇 曹荣刚[1,2,3] 李发栋[1] 曹宇曦 李佳武 张安琪 LI Ping;ZHOU Yu;CAO Ronggang;LI Fadong;CAO Yuxi;LI Jiawu;ZHANG Anqi(School of Mechatronical Engineering,Beijing Institute of Technology,Beijing 100081,China;Science and Technology on Electromechanical Dynamic Control Laboratory,Beijing Institute of Technology,Beijing 100081,China;Tangshan Research Institute,Beijing Institute of Technology,Tangshan 063611,Hebei,China)

机构地区：[1]北京理工大学机电学院,北京100081 [2]北京理工大学机电动态控制重点实验室,北京100081 [3]北京理工大学唐山研究院,河北唐山063611

出　　处：《兵工学报》2024年第6期1747-1760,共14页Acta Armamentarii

摘　　要：红外成像制导系统受到严苛运行环境影响,其成像过程伴随复杂背景噪声干扰,严重影响系统制导跟踪精度。为减少复合噪声对红外成像效果的影响,在分析多种常见噪声的成因和特性的基础上,提出一种基于加性成分和乘性成分的噪声特性先验设定,结合空间域和变换域的双域融合去噪思想,设计了一种基于深度卷积神经网络的多类型噪声去除方法。该方法将富梯度流卷积模块引入UNet++结构以缩减梯度信息冗余并提升多感受野特征提取能力;针对噪声形态特性提出维度注意力机制以实现双域噪声估计;引入高阶双树复小波变换作为域变换方法,提升在不同尺度和方向上对噪声成分的识别能力。通过消融实验验证了噪声先验设定以及双域融合去噪思想的有效性和优越性,通过对比实验证明了所提方法对多种类型噪声均具有优秀的去噪能力。所提方法对高斯噪声的去噪峰值信噪比和结构相似度指标分别达到29.57和0.85,优于其他典型噪声抑制方法;对多类型混合噪声则分别达到27.84和0.82,达到良好的去噪水平。此外,也验证了所提方法对真实图像噪声具有优秀的去噪能力。The infrared imaging guidance system is severely affected by harsh operating environments,and the imaging process is accompanied by complex background noise interference,which seriously affects the guidance and tracking accuracy of the system.In order to reduce the impact of composite noise on the infrared imaging effect,a priori setting of noise characteristics based on additive and multiplicative components is proposed based on the analysis of the causes and characteristics of various common noises;further,a denoising method for different types of noise based on deep convolutional neural network is designed according to the dual-domain fusion denoising idea of spatial domain and transform domain.This method introduces a rich gradient flow convolutional module into the UNet++structure to reduce the gradient information redundancy and enhance the multi-receptive field feature extraction ability.A dimension attention mechanism is proposed to achieve dual-domain noise estimation according to the noise morphology characteristics.The high-order dual-tree complex wavelet transform is introduced as the domain transformation method to improve the recognition ability of noise components at different scales and directions.The effectiveness and superiority of the noise prior setting and dual-domain fusion denoising idea are verified through ablation experiments,and the proposed method demonstrates the excellent denoising ability for various types of noise through comparative experiments.The proposed method achieves the peak signal-to-noise ratio of 29.57 and the structural similarity index of 0.85 for Gaussian noise removal,which is superior to other typical noise suppression methods.For multi-type mixed noise,it achieves the denoising levels of 27.84 and 0.82,respectively.Moreover,the proposed method was validated to possess significant capability in removing the noises from real image.

关 键 词：红外成像 图像去噪 深度学习 空间域去噪 变换域去噪 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]