微聚焦X射线图像自适应复原关键理论与技术研究

This project focuses on the research of self-adaptive micro-focus X-ray image restoration, which is application-driven by the high-speed and high-precision requirements of micro-focus X-ray imaging in integrated circuit (IC) packaging process. .Firstly, a noise-adaptive fidelity term model for mixed noise is established based on different statistical information, result in overcoming the drawbacks of single fidelity norm which is just useful for a certain type of noise. .Secondly, an equivalent noise-adaptive fidelity item under mixed noise corrupted will be established in the derivative space, which is based on the study of transformation between image space and related derivative space. Due to the characterization of gradient features in derivative space, the capacity of detail-preserving will be improved. .Thirdly, a Riemann manifold regularization model will be proposed for high-precision image restoration, especially for IC micro details under high-speed X-ray imaging. In order to obtain more accurate descriptions of the image prior knowledge on Riemann manifolds, the methods of differential geometry will be employed to determine the classification standard of local geometric features, while the particle swarm optimization algorithms will be used for accurate classification..Finally, fast and effective algorithms will be studied, including adaptive parameter adjustment scheme will be designed to balance the effectiveness of noise-removing and detail-preserving. To avoid the tendency to plunge into local optimizations caused by improper initial value and noise, the global best harmony search algorithm will be used to obtain the initial value of iteration. A spectrum adaptive method based on multi-resolution decomposition will be introduced to obtain the superiority of computational speed.

本项目从集成电路微聚焦X射线成像的高速高精度需求出发，开展应用驱动的微聚焦X射线图像自适应复原理论与技术研究。首先，针对不同噪声统计特性，建立噪声自适应保真项模型，突破单一保真项仅去除单一噪声的弊端；其次，研究混合噪声条件下图像微分空间与图像空间的转化条件，建立微分空间内噪声自适应保真项的等效模型，有效利用微分空间对于图像梯度特征的刻画来实现细节特征的保持；第三，建立基于几何特征分类的黎曼流形正则化模型。利用微分几何工具确定图像局部几何特征的分类标准，并采用粒子群随机搜索精确分类，从而在流形上获得更精确的图像先验知识描述，实现高速运动成像条件下集成电路细小几何特征的高可靠性复原；最后，研究快速有效的求解算法，包括：自适应参数调整方案，解决噪声去除和细节保持的平衡问题；全局和声搜索确定迭代初始值，解决初值选取不当和噪声引起的局部最优问题；采用多分辨率谱自适应方法，解决算法求解的速度问题。

本项目以集成电路制造过程中的微聚焦X射线成像检测为应用背景，以影响精度和速度的共性基础理论难题——微聚焦X射线图像复原技术为研究对象，完成了应用驱动的微聚焦X射线图像复原理论研究。主要内容包括：1）完成了微聚焦X射线的成像机理及集成电路X射线图像特点的分析，针对振动引起的运动模糊完成了两种鲁棒自适应振动抑制方法的设计；2）针对运动模糊、噪声强烈、灰度水平低、对比度低的图像特点，提出了正则化框架下基于伪逆矩阵保真项的全变分正则化图像复原模型及算法、微分空间下的广义总变分正则化图像复原模型及算法、针对柔性PCB板的流形正则化图像复原方法，以完成图像的有效复原；3）完成了基于图像灰度特征的分区方法及基于分区的多正则化图像复原算法的设计，对细节区域和平滑区域分别选取先验知识，解决了噪声去除和细节保持的平衡问题；4）完成了一种新的飞蛾火焰优化算法、基于多分辨率和和声搜索的自适应复原方法、一种自适应参数的动量梯度投影算法的设计，提升了全局搜索能力和算法的收敛速度；5）拓展研究了集成电路的内部缺陷检测方法和特点。实验结果表明，以上模型和算法均可获得更优的图像复原效果，达到了细节保持和噪声去除的有效平衡，可为后续自动化和智能缺陷检测提供原始干净的有效图像，进而提升检测精度。本项目解决的微聚焦X射线图像有效复原问题，是微聚焦X射线成像检测、计算机断层影像分析的共性技术难题，研究成果不仅可为集成电路高端芯片封装过程的超精密X射线成像检测提供理论基础，还可提高我国集成电路封装测试技术水平及微聚焦X射线检测装备的自主研发能力。同时，正则化框架下的图像复原模型及快速精确的求解算法，可为遥感成像、人脸识别、光流估计、图像去雾等应用领域的正则化方法提供理论支撑。