医学影像处理的深度学习可解释性研究进展

随着医学影像数据的迅速增长，传统的影像分析方法给医生带来巨大挑战。利用计算机视觉技术提供自动或半自动辅助诊断，可大大缓解人工阅片压力，提高诊断的准确性，促进医疗流程的标准化建设等。目前，深度学习卷积神经网络在医学影像处理中已取得不俗表现，但深度学习“黑匣子”的不可解释性阻碍了智能医疗诊断的发展。为增强对医学影像数据处理的深度学习可解释性的了解，对近几年相关研究进展进行了综述。首先，综述了深度学习在医学领域的应用现状及面临的问题，对神经网络的可解释性内涵进行了讨论；然后，从现有深度学习可解释性的常见方法出发，重点讨论了医学影像处理的深度学习可解释性研究进展；最后，探讨了医学影像处理的深度学习可解释性的发展趋势。     

Effect of the nonlinearity of the CCD in Fourier transform profilometry on spectrum overlapping and measurement accuracy

In Fourier transform profilometry (FTP), we must restrain spectrum overlapping caused by the nonlinearity of the charge coupled device (CCD) and increase the measurement accuracy of the object shape. Firstly, the causes of producing higher-order spectrum components and inducing spectrum overlapping are analysed theoretically, and a simple physical explanation and analytical deduction are given. Secondly, aiming to suppress spectrum overlapping and improve measurement accuracy, the influence of spatial carrier frequency of projection grating on them is analysed. A method of increasing the spatial carrier frequency of projection grating to restrain or reduce the spectrum overlapping significantly is proposed. We then analyze the mechanism of how the spectrum overlapping is reduced. Finally, the simulation results and experimental measurements verify the correction of the proposed theory and method.     

面向自然场景图像的三阶段文字识别框架

文字识别技术在文档管理、图像理解、视觉导航等中具有重要应用。然而，自然场景中的文字通常排列任意、形状不一、字体多样，难以被检测和识别。提出了面向自然场景图像的三阶段文字识别框架，该框架包括文字检测、文字矫正和文字识别。首先，利用特征金字塔网络分割图像中的字符，基于双向长短期记忆网络获取字符间的亲和度，连接孤立字符构建单词行，文字检测率（F分数）高达91.97%。然后，通过多目标矫正网络矫正被检测文字，以应对场景图像文字的复杂形变，增强阅读性。最后，通过注意力序列识别网络按序输出预测结果，实现单词级识别，文字识别正确率达84.98%。     

一种方向梯度直方图的降维方法

为描述对象的局部外观和形状,方向梯度直方图首先将图像划分成小区域(被称为cell),然后在其上累加像素梯度方向的一维直方图.在被称为block的较大区域(由数个相邻的cell组成)上连接cell的直方图,经归一化处理形成特征向量.为减弱由block引起的区域量化走样,在计算检测窗口的特征时,采取部分重叠block的措施,从而大大增加了特征维度以及目标检测时的计算量.通过扩大参与相邻cell之间像素梯度插值的面积,并设置适当的高斯平滑核尺度,可消除block重叠,从而将64×128尺寸的窗口的方向梯度直方图特征维度由3 780降低为1 152.INRIA的行人数据集实验表明,该方法也可减弱区域量化走样,且其性能与原方向梯度直方图几乎相当,而检测速度却显著提高.     

Effect of the nonlinearity of CCD in Fourier transform profilometry on spectrum overlapping and measurement accuracy

In Fourier transform profilometry (FTP), we must restrain spectrum overlapping caused by the nonlinearity of charge coupled device (CCD) and increase the measurement accuracy of object shape. Firstly, the causes of producing higher-order spectrum components and inducing spectrum overlapping are analysed theoretically, and simple physical explanation and analytical deduction are given. Secondly, aiming to suppress spectrum overlapping and improve measurement accuracy, the influence of spatial carrier frequency of projection grating on them is analysed. A method of increasing the spatial carrier frequency of projection grating to restrain or reduce the spectrum overlapping significantly is proposed. We then analyze the mechanism of how the spectrum overlapping is reduced. Finally, the simulation results and experimental measurements verify the correction of the theory and method proposed.