红外热成像信号特征提取和降噪理论及其方法

The temperature field provides essential information of operation of high-end equipments, feature extraction on three dimensional temperature profile could be effectively applied to identify working status of the equipments. Since infrared thermal imaging technology is the most commonly used and most effective solution to obtain real-time and accurate temperature information, so it is important to systematically investigate the characteristics of infrared image noise signal. It is observed that captured infrared images contain significant amount of fixed pattern noise (FPN) and high frequency strip noise. In our research, we firstly plan to develop effective and shutter-less denoising algorithms to eliminate such noises without removing valuable infrared image fine details, so to improve the performance of infrared thermal imaging for accurate and uninterrupted scene monitoring. Research on the FPGA-based parallel computing is also performed to ensure the real-time hardware implementation of the proposed signal denoising algorithms. Then we develop data mining solutions, based on image detail enhancement and dynamic tone mapping, to discover local insignificant temperature variances within the captured high dynamic range (HDR) infrared data. Finally, we propose solutions for pixel-wise visual and infrared image registration based on edge extraction and matching. The established pixel-to-pixel correspondences enable us to perform effective fusion of infrared and visible images, and further to reconstruct 3D real-time temperature profile of the target object which can be used for accurate detection and localization of abnormal temperature variances of high-end equipments.

温度场包含丰富的高端装备运行状态特征，通过三维温度场的特征提取，可识别设备的工作运行状态，而红外热成像是实时、准确获得温度场信息的有效手段，为此，探索红外热成像信号及噪声的形成规律和存在形态，针对红外成像特有的固定模式噪声和高频线状动态信号噪声，提出相应的适用于红外图像的信号除噪方法，在不损失红外图像宝贵细节信息的前提下有效降低图像噪声干扰，提升红外热成像的成像精度和无间断监测性能。研发基于硬件FPGA的并行运算算法，确保信号除噪的实时性和硬件可实现性。在此基础上，通过细节增强和动态数据映射等信号处理技术，挖掘隐藏在高动态范围红外数据里的局部微小温度变化；提出基于轮廓边缘的像素级别红外图像和可见光图像匹配算法，通过红外图像与可见光图像的有效融合，实现三维温度场的重建，从而将红外热成像应用于高端装备的三维立体全方位覆盖实时监测，对出现的异常温度变化区域进行精确定位和实时报警。

项目针对红外热成像存在固定模式噪声、低分辨率、信息单一（只有温度信息）等技术瓶颈问题，开展了噪声机理、图像质量增强和三维温度场重建的探索性研究，通过红外图像的数据优化、数据增强、和数据融合三个方面的攻关，提出一整套适用于红外图像信号的特征提取、降噪、增强和融合的理论及方法，提升红外热成像的精确性、可用性和多维性。系统地研究了红外热成像噪声信号的形成规律和存在形态，提出了红外成像无挡板校正方法，提出了基于卷积神经网络的条状噪声消除方法，提升红外热成像的成像精度和无间断监测性能。针对红外图像分辨率低，构建了多感受野级联深度神经网络，提出了适用于红外图像的超分辨率算法，从算法层面提升红外热成像的图像质量，挖掘局部微小温度变化；为了更好的利用温度信息，提出红外热成像的低纹理图像与可见光相机采集的高纹理图像的像素级别匹配方法，通过红外图像、可见光图像与深度信息的有效融合，实现三维温度场的重建，为红外热成像应用于高端装备的三维立体全方位覆盖实时监测提供技术支撑。