.960· 工程科学学报，第41卷，第7期 佳的图像配准效果 (潘博，贲进，闫佳雯，等.全局约束下超声图像微器械轮廓 为进一步验证本文算法的准确性，采用峰值信 提取方法.哈尔滨工业大学学报，2018,50(1)：24) 噪比(PSNR)与归一化互信息(NMI)对上述图像纠 [3]Dellinger F,Delon J,Gousseau Y,et al.SAR-SIFT:a SIFT- like algorithm for SAR images.IEEE Trans Geosci Remote Sens, 正结果与原始参考图像的差值进行客观评估，结果 2015,53(1):453 如表1和表2所示. [4] Sariyanidi E,Gunes H,Cavallaro A.Automatic analysis of facial 表1不同算法的峰值信噪比对比结果 affect:a survey of registration,representation,and recognition Table 1 Comparison of PSNR values of different methods dB IEEE Trans Pattern Anal Mach Intell,2015,37(6):1113 [5]Thirion J P.Image matching as a diffusion process:an analogy 不同算法 图像 with Maxwell's demons.Med Image Anal,1998,2(3):243 H-S Brox SIFT Flow本文算法 [6]Horn B K P,Schunck B G.Determining optical flow.Artif Intell, 核磁共振 13.30 16.99 16.89 18.09 1981,17(1-3):185 柔性 18.69 21.57 25.23 25.24 [7]Nagel HH,Enkelmann W.An investigation of smoothness con- 人脸 20.27 20.35 22.08 22.82 straints for the estimation of displacement vector fields from image sequences.IEEE Trans Pattern Anal Mach Intell,1986,8(5):565 表2不同算法的归一化互信息对比结果 [8]Brox T.Malik J.Large displacement optical flow:descriptor Table 2 Comparison of NMI values of different methods matching in variational motion estimation.IEEE Trans Pattern A- nal Mach Intell,2011,33(3):500 不同算法 图像 [9] Amiaz T,Lubetzky E,Kirvati N.Coarse to over-fine optical flow H-S Brox SIFT Flow本文算法 estimation.Pattern Recognit,2007,40(9):2496 核磁共振 1.0279 1.1031 1.1255 1.1545 [10]Bao LC,Yang QX,Jin H L.Fast edge-preserving patch match 柔性 1.5076 1.5027 1.1790 1.5216 for large displacement optical flow /Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Colum- 人脸 1.4180 1.4324 1.4060 1.4554 bus,2014:3534 从表中可以看出：本文算法在峰值信噪比与归 [11]Liu C.Yuen J,Torralba A.SIFT flow:dense correspondence 一化互信息这两项评价指标上，数值结果均高于传 across scenes and its applications.IEEE Trans Pattern Anal Mach Intell,2011,33(5):978 统算法，表明本文算法取得了最优的配准效果 [12]Sun D Q,Roth S,Black M J.A quantitative analysis of current 4结论 practices in optical flow estimation and the principles behind them.Int J Comput Vision,2014,106(2):115 (1)针对传统光流模型在非刚性图像配准中存 [13]Liu H B,Chang F L.Moving object detection by optical flow 在的问题，提出利用各向异性正则项，根据梯度进行 method based on adaptive weight coefficient.Opt Precis Eng, 自适应平滑，保护了图像的边缘特征：此外，通过在 2016,24(2):460 (刘红彬，常发亮.权重系数自适应光流法运动目标检测. 光流场模型中引入一个非局部平滑项，可充分利用 光学精密工程，2016,24(2)：460) 邻域信息来正则化光流矢量场，通过自适应邻域加 [14] Revaud J,Weinzaepfel P,Harchaoui Z,et al.EpicFlow:edge- 权，既达到了去除光流噪点的目的，又较好地保持了 preserving interpolation of correspondences for optical flow/ 图像的细节特征，提高了图像配准精度 Proceedings of the IEEE International Conference on Computer Vi- (2)利用交替求解的方法计算位移场，用最终 sion and Pattern Recognition.Boston,2015:1164 [15]Li S,Fan X G.Xu Y L,et al.Bio-inspired motion-adaptive es- 位移场矫正浮动图像实现了图像配准.与传统算法 timation algorithm of sequence image.Chin J Eng,2017,39 的仿真实验对比结果表明：本文算法具有较高的鲁 (8):1238 棒性，能够获得更理想的非刚性图像配准效果. (李帅，樊晓光，许悦雷，等。序列图像运动自适应V1-T 光流估计算法.工程科学学报，2017,39(8)：1238) 参考文献 [16]Dosovitskiy A,Fischer P,Ilg E,et al.FlowNet:learning optical [1]Chen S W,Zhang S X,Yang X G,et al.Registration of visual- flow with convolutional networks /Proceedings of the IEEE In- infrared images based on ellipse symmetrical orientation moment. ternational Conference on Computer Vision.Santiago,2015:2758 Chin J Eng,2017,39(7):1107 [17]He K,Yan JX,Wei Y,et al.Non-rigid image registration using (陈世伟，张胜修，杨小冈，等.基于椭圆对称方向矩的可见 improved optical flow field.J Tianjin Unin Sci Technol,2018,51 光与红外图像配准算法.工程科学学报，2017.39(7)：1107) (5):491 [2]Pan B,Ben J,Yan J W,et al.Method to extract micro device (何凯，闫佳星，魏颖，等。基于改进光流场模型的非刚性图 profile in ultrasound image under global constraints.Harbin Inst 像配准.天津大学学报：自然科学与工程技术版，2018,51 Technol,2018,50(1):24 (5):491)工程科学学报,第 41 卷,第 7 期 佳的图像配准效果. 为进一步验证本文算法的准确性,采用峰值信 噪比(PSNR)与归一化互信息(NMI)对上述图像纠 正结果与原始参考图像的差值进行客观评估,结果 如表 1 和表 2 所示. 表 1 不同算法的峰值信噪比对比结果 Table 1 Comparison of PSNR values of different methods dB 图像 不同算法 H鄄鄄 S Brox SIFT Flow 本文算法 核磁共振 13郾 30 16郾 99 16郾 89 18郾 09 柔性 18郾 69 21郾 57 25郾 23 25郾 24 人脸 20郾 27 20郾 35 22郾 08 22郾 82 表 2 不同算法的归一化互信息对比结果 Table 2 Comparison of NMI values of different methods 图像 不同算法 H鄄鄄 S Brox SIFT Flow 本文算法 核磁共振 1郾 0279 1郾 1031 1郾 1255 1郾 1545 柔性 1郾 5076 1郾 5027 1郾 1790 1郾 5216 人脸 1郾 4180 1郾 4324 1郾 4060 1郾 4554 从表中可以看出:本文算法在峰值信噪比与归 一化互信息这两项评价指标上,数值结果均高于传 统算法,表明本文算法取得了最优的配准效果. 4 结论 (1)针对传统光流模型在非刚性图像配准中存 在的问题,提出利用各向异性正则项,根据梯度进行 自适应平滑,保护了图像的边缘特征;此外,通过在 光流场模型中引入一个非局部平滑项,可充分利用 邻域信息来正则化光流矢量场,通过自适应邻域加 权,既达到了去除光流噪点的目的,又较好地保持了 图像的细节特征,提高了图像配准精度. (2)利用交替求解的方法计算位移场,用最终 位移场矫正浮动图像实现了图像配准. 与传统算法 的仿真实验对比结果表明:本文算法具有较高的鲁 棒性,能够获得更理想的非刚性图像配准效果. 参 考 文 献 [1] Chen S W, Zhang S X, Yang X G, et al. Registration of visual鄄 infrared images based on ellipse symmetrical orientation moment. Chin J Eng, 2017, 39(7): 1107 (陈世伟, 张胜修, 杨小冈, 等. 基于椭圆对称方向矩的可见 光与红外图像配准算法. 工程科学学报, 2017, 39(7): 1107) [2] Pan B, Ben J, Yan J W, et al. Method to extract micro device profile in ultrasound image under global constraints. J Harbin Inst Technol, 2018, 50(1): 24 (潘博, 贲进, 闫佳雯, 等. 全局约束下超声图像微器械轮廓 提取方法. 哈尔滨工业大学学报, 2018, 50(1): 24) [3] Dellinger F, Delon J, Gousseau Y, et al. SAR鄄鄄 SIFT: a SIFT鄄鄄 like algorithm for SAR images. IEEE Trans Geosci Remote Sens, 2015, 53(1): 453 [4] Sariyanidi E, Gunes H, Cavallaro A. Automatic analysis of facial affect: a survey of registration, representation, and recognition. IEEE Trans Pattern Anal Mach Intell, 2015, 37(6): 1113 [5] Thirion J P. Image matching as a diffusion process: an analogy with Maxwell蒺s demons. Med Image Anal, 1998, 2(3): 243 [6] Horn B K P, Schunck B G. Determining optical flow. Artif Intell, 1981, 17(1鄄3): 185 [7] Nagel H H, Enkelmann W. An investigation of smoothness con鄄 straints for the estimation of displacement vector fields from image sequences. IEEE Trans Pattern Anal Mach Intell, 1986, 8(5): 565 [8] Brox T, Malik J. Large displacement optical flow: descriptor matching in variational motion estimation. IEEE Trans Pattern A鄄 nal Mach Intell, 2011, 33(3): 500 [9] Amiaz T, Lubetzky E, Kiryati N. Coarse to over鄄fine optical flow estimation. Pattern Recognit, 2007, 40(9): 2496 [10] Bao L C, Yang Q X, Jin H L. Fast edge鄄preserving patch match for large displacement optical flow / / Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Colum鄄 bus, 2014: 3534 [11] Liu C, Yuen J, Torralba A. SIFT flow: dense correspondence across scenes and its applications. IEEE Trans Pattern Anal Mach Intell, 2011, 33(5): 978 [12] Sun D Q, Roth S, Black M J. A quantitative analysis of current practices in optical flow estimation and the principles behind them. Int J Comput Vision, 2014, 106(2): 115 [13] Liu H B, Chang F L. Moving object detection by optical flow method based on adaptive weight coefficient. Opt Precis Eng, 2016, 24(2): 460 (刘红彬, 常发亮. 权重系数自适应光流法运动目标检测. 光学精密工程, 2016, 24(2): 460) [14] Revaud J, Weinzaepfel P, Harchaoui Z, et al. EpicFlow: edge鄄 preserving interpolation of correspondences for optical flow / / Proceedings of the IEEE International Conference on Computer Vi鄄 sion and Pattern Recognition. Boston, 2015: 1164 [15] Li S, Fan X G, Xu Y L, et al. Bio鄄inspired motion鄄adaptive es鄄 timation algorithm of sequence image. Chin J Eng, 2017, 39 (8): 1238 (李帅, 樊晓光, 许悦雷, 等. 序列图像运动自适应 V1鄄鄄 MT 光流估计算法. 工程科学学报, 2017, 39(8): 1238) [16] Dosovitskiy A, Fischer P, Ilg E, et al. FlowNet: learning optical flow with convolutional networks / / Proceedings of the IEEE In鄄 ternational Conference on Computer Vision. Santiago, 2015: 2758 [17] He K, Yan J X, Wei Y, et al. Non鄄rigid image registration using improved optical flow field. J Tianjin Univ Sci Technol, 2018, 51 (5): 491 (何凯, 闫佳星, 魏颖, 等. 基于改进光流场模型的非刚性图 像配准. 天津大学学报:自然科学与工程技术版, 2018, 51 (5): 491) ·960·