动量加速方法提高层叠衍射成像的质量  

Improving the reconstructed image quality in ptychography with momentum acceleration

作　　者：陶旭磊 许子健[1,2] 王春鹏 邰仁忠[1,2] TAO Xulei;XU Zijian;WANG Chunpeng;TAI Renzhong(Shanghai Institute of Applied Physics, Chinese Academy of Sciences, Shanghai 201800, China;Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai 201210, China;University of Chinese Academy of Sciences, Beijing 100049, China)

机构地区：[1]中国科学院上海应用物理研究所,上海201800 [2]中国科学院上海高等研究院,上海201210 [3]中国科学院大学,北京100049

出　　处：《核技术》2019年第5期10-16,共7页Nuclear Techniques

基　　金：国家自然科学基金(No.11575283;No.11875316;No.U1832146);国家重点研发计划(No.2017YFA0403400)资助~~

摘　　要：X射线层叠衍射成像技术(Ptychography)所采用的图像重建迭代算法是一种传统的梯度下降算法。该算法从一组猜测函数开始,实验中探测到的正确信息被不断代入循环迭代中来修正猜测函数,最终达到收敛。这种算法存在一个问题就是容易出现收敛停滞的现象,也就是所谓的陷入局域最小值。造成这种现象出现的原因是局域最小值附近的梯度极小导致梯度下降的幅度极小。为了解决这一问题,最近"动量"这一概念被引入Ptychography算法中,产生了动量加速层叠衍射迭代引擎(momentum Ptychographical Iterative Engine,mPIE)算法。加入"动量"相当于给迭代重建过程赋予了"惯性"性质,当前梯度下降的幅度不仅由当前梯度决定,还会受到之前变化的幅度影响。所以即便是在梯度极小的局域最小值附近,目标函数仍然可以跳出停滞,继续变小。为了验证"动量"的引入对图像重建的提升效果,详细比较了动量mPIE算法和传统的拓展层叠衍射迭代引擎(extending Ptychographical Iterative Engine,e PIE)算法对同一套扫描相干衍射成像实验数据的重建图像质量。重建结果表明:相对于传统重建算法,动量加速算法可以获得更好的图像重建质量。[Background] The image retrieval algorithm of ptychography is a kind of conventional gradient descent algorithm. This algorithm begins with a set of guessing functions, and the signals collected in the experiment are continuously input into the iteration process to achieve final convergence. However the reconstructed process often falls into stagnation because of the low gradient around the local minima and the extent of the gradient descent is too small.[Purpose] This study aims to solve the convergence stagnation problem of conventional gradient descent algorithm by using momentum ptychographical iterative engine (mPIE).[Methods] The concept of "momentum" was introduced into the ptychography algorithm to update conventional algorithm, called the momentum ptychographical iterative engine (mPIE). Adding momentum to an algorithm means that the iterative process has the characteristic of inertance. The extent of gradient descent relies on the previous steps and the gradient of current step. With the momentum, convergence can escape stagnation, and the reconstructed process can continue even around the local minima with a gradient is very small. The same set of coherent diffraction imaging experimental data was used to test the reconstructed image quality of mPIE algorithm and compare with that of extending ptychographical iterative engine (ePIE) algorithm.[Results] The experimental results showed that mPIE algorithm outperformed ePIE algorithm with smaller convergence value of objective function, and better final reconstructed image quality.[Conclusion] The result indicates that a better reconstructed result can be achieved with mPIE algorithm compared with the traditional ePIE algorithm.

关 键 词：X射线扫描相干衍射成像 动量 X射线显微 相位恢复迭代算法 

分 类 号：TL99[核科学技术—核技术及应用]