9期张秀伟等: 基于 Co-motion 的可见光热红外图像序列自动配准算法 (a) 建筑物场景正午时的红外可见光图像对 (a) Thermal-visual image pair of building scene at noon (b) 道路场景红外可见光图像对 (b) Thermal-

Size: px

Start display at page:

Download "9期张秀伟等: 基于 Co-motion 的可见光热红外图像序列自动配准算法 (a) 建筑物场景正午时的红外可见光图像对 (a) Thermal-visual image pair of building scene at noon (b) 道路场景红外可见光图像对 (b) Thermal-"

釜隗
7 years ago
Views:

1 36 9 Vol. 36, No ACTA AUTOMATICA SINICA September, 2010 Co -motion Co-motion, Co-motion,.,,,.,,. 8. DOI, Co-motion,, /SP.J Automatic Visual-thermal Image Sequence Registration Based on Co-motion ZHANG Xiu-Wei 1 ZHANG Yan-Ning 1 YANG Tao 1 ZHANG Xin-Gong 1 SHAO Da-Pei 1 Abstract An automatic visual thermal image sequence registration method based on co-motion was proposed. Different from other methods, co-motion (concurrent motions) statistics feature was adopted to regist heterogeneous image sequences. Compared with registration based on co-motion between homogeneous image sequences, our method faced many difficulties. By analyzing the difficulties, we proposed the corresponding point pairs extraction method and outliers removal method, which are suitable to visual thermal image sequence. We demonstrated the performance of the method on eight groups of visual thermal synchronous video sequences, and the results showed that the proposed algorithm carried out precise image registration under the change of image rotation, translation, scaling and viewing angle. Experiments demonstrated the accuracy and robustness of the proposed method. Key words Image registration, concurrent motion (co-motion), visual image sequence, thermal image sequence, [1 2].,.., Manuscript received March 17, 2009; accepted March 8, 2010 (863 ) (2009AA01Z315), ( , ), (708085), ( ) Supported by National High Technology Research and Development Program (863 Program) of China (2009AA01Z315), National Natural Science Foundation of China ( , ), Cultivation Fund of the Key Scientific and Technical Innovation Project, Ministry of Education of China (708085), and China Postdoctoral Science Foundation ( ) Shannxi Provincial Key Laboratory of Speech and Image Information Processing, College of Computer Science, Northwest Polytechnical University, Xi an : 1) : 1,,. 2) :,.,,, 1 (b), 1 (b) ( ). ( 1 (b), ),. 3) :,.,. 4) :,. 1 (a) 1 (c) 1 (d)

Co -motion 1 1 1 1 1 Co-motion, Co-motion,.,,,.,,. 8. DOI, Co-motion,, 10.3724/SP.J.1004.2010.

2 9期张秀伟等: 基于 Co-motion 的可见光热红外图像序列自动配准算法 (a) 建筑物场景正午时的红外可见光图像对 (a) Thermal-visual image pair of building scene at noon (b) 道路场景红外可见光图像对 (b) Thermal-visual image pair of path scene (c) 建筑物场景傍晚时的红外可见光图像对 (c) Thermal-visual image pair of building scene at afternoon (d) 建筑物场景夜晚时的红外可见光图像对 (d) Thermal-visual image pair of building scene at night 图1 Fig. 1 热红外可见光图像配准示例 Examples of thermal visual image pair 场景拍摄的三组图像, 它们间的灰度变化十分明显. 与单类型传感器相比, 从红外和可见光图像中提取一致性图像特征的困难更大. 因此基于静态图像特征的异源配准方法很难获得鲁棒的配准结果, 它主要包括基于像素灰度的配准[3 4] 基于图像几何特征的配准[5 7] 和基于图像分类的配准[8]. 基于运动特征的配准是近年来出现的一类颇受关注的方法, 该类方法利用图像序列中的运动信息 (运动目标的侧影轮廓轨迹等) 进行配准. 在同源图像序列配准中该方法使用的较多, 它们大多需要精确的运动轨迹或运动检测结果[9 10]. Ju 等[11] 也尝试利用运动信息来配准可见光热红外图像序列, 在仅允许一个目标在室内场景中运动的情况下, 通过提取准确的运动目标轮廓进行可见光热红外 1221 图像序列的配准. 由可见光热红外图像配准面临的问题可知, 在无任何约束 (如背景和运动都很复杂的室外监控场景) 情况下, 同时从可见光序列和热红外序列中获得精确的运动信息是十分困难的. 由 Szlavik 等提出的基于 Co-motion 的可见光序列配准[12 16] 为运动特征配准注入了新鲜血液, 它是一类不依赖精确运动检测的配准方法, 通过对具有部分重叠视野的同步视频的运动检测, 统计两个序列中像素点同时发生运动的情况, 并依据该统计信息选择同名点对. 基于 Co-motion 的配准方法为可见光热红外图像序列自动配准提供了可行的参考. 但是由于成像机理不同, 将 Co-motion 特征运用于可见光热红外图像序列配准仍面临一些问题: 1) 要求观测场景中具有可视的运动的热目标. 它是确保两序列在重叠区域中提取的前景具有交集的前提. 这一点较易满足, 因为在大多数监控场景中, 被观测的对象多为行人和汽车. 2) 两同源图像序列在重叠区域内的前景背景对比度具有相似性, 因此两图像序列在重叠区域内的运动检测结果较为一致, 基于检测结果的运动统计特征也呈现出较强的相似性. 然而可见光和热红外具有不同的成像特性, 在重叠区域内两者的前景背景对比度差异较大, 如何保证两图像序列在重叠区域内的运动前景具有尽量大的交集是影响运动统计特征使用的关键问题. 3) 文献 [12 16] 没有给出大尺度变化下配准结果, 也没有针对该问题进行分析和说明. 相对于同源系统, 异源系统更易面临传感器间尺度变化较大的问题. 这是由于热红外传感器较为昂贵, 异源系统更倾向于用热红外进行大范围观测和定位, 用可见光传感器获得精细观测信息. 4) 相对于同源序列配准, 不同的成像特性加之红外成像对比度低边缘模糊, 使得不同源配准面临大量的外点干扰, 文献 [12 16] 等提出的外点去除方法并不适宜. 本文通过引入 Co-motion 特征来解决异类传感器图像序列配准问题. 由于同源配准和异源配准存在很大差异, 本文针对在异源配准应用 Comotion 所面临的上述难题, 提出了适宜于异源序列的特征提取方法和外点剔除方法. 最后使用 LM (Levenberg-Marquardt) 算法[17 18] 迭代优化模型参数. 实验结果表明, 本文算法在大尺度旋转及平移变化下均能精确配准. 1 系统框架图 2 给出了可见光热红外图像序列配准系统的流程图, 它是以可见光热红外同步视频为输入,

1 热红外可见光图像配准示例 Examples of thermal visual image pair 场景拍摄的三组图像, 它们间的灰度变化十分明显. 与单类型传感器相比, 从红外和可见光图像中提取一致性图像特征的困难更大. 因此基于静态图像特征的异源配准方法很难获得鲁棒的配准结果, 它主要包括基于像素灰度的配准[3 4] 基于图像几何特征的配准[5 7] 和基于图像分类的配准[8].

3 Co-motion Fig. 2 Flow chart of the proposed method,.,. : 1),,.,..,,. 2),, / (Random sample consensus, RANSAC) [19]. 3) 2),, LM. 2 Co-motion 2.1 Co-motion 1) 1, Co-motion. [20],. [21]..., seq(i, j), (1) ( ), (i, j), [16]., A, B, N A M A A, N B M B B, T. 2) [14 15].,,.,,,,.,,,.,,,., (2) ( ), validateseq(l, i, j) true, L (i, j) ;., v max, v min, low thd hgh thd.

4 9期张秀伟等: 基于 Co-motion 的可见光热红外图像序列自动配准算法 seql (i, j) = [seq ( L (i, j, 1),, seql (i, j, t),, seql (i, j, T )] 0, (i, j) detected as background at t-frame seql (i, j, t) = 1, (i, j) detected as foreground at t-frame 1 i NL, 1 j ML, L {A, B} µt T P P validateseq(l, i, j) = seql (i, j, t) > vmin seql (i, j, t) < vmax, L [A, B] t=1 t=1 µt P vm AX = max seql (i, j, t) 1223 (1) µ 1<i<N,1<j<M,L [A,B] t=1 (2) vmin = vm AX low thd vmax = vm AX hgh thd seq (i, j, t) seq (p, q, t) A B t=1 mapa(i, j) = min T P (p,q) U (i,j) V seqa (i, j, t) T P (3) t=1 3) 候选同名点对选取运动状态序列是二值序列, 海明距离是度量这类序列相似性的一种有效方法. 考虑到运动序列间的距离还与运动状态序列 (1) 中的总数相关, 本文采用了如式 (3) 所示的相似性度量. 对基准集合 A 中的有效运动状态序列, 从待配准集合 B 的有效运动状态序列中, 选择与其距离最近的像素点作为匹配点, 共同形成候选同名点对. 函数 mapa(i, j) 求取的就是与基准 A 中某像素点运动状态最相似的 B 像素集合. 其中, V 为图像 A 满足式 (2) 的像素点集合; U 为图像 B 满足式 (2) 的像素点集合. 从图 3 (b) 中可以看出, 经多对一映射去除后, 所剩下的同名点对大多为错误点对. (a) 候选同名点对 (a) Original corresponding point pairs 2.2 外点去除 1) 多对一映射的处理基于 Co-motion 生成的候选匹配点对中存在两种多对一映射, 即一个基准点对应多个最优候选匹配点的映射和多个基准点对应同一个匹配点的映射. 在这些多对一映射中, 有些是由于尺度差异造成的, 属于正确的配准点对, 还有些是由于运动检测不准确或者非重叠区域的干扰点造成的, 属于错误的配准点对, 正确与错误的点对很难区分, 文献 [12 13, 16] 采取了直接剔除多对一映射的点对. 随着基准和待配准图像序列间尺度差异的增大, 由尺度差异引起的多对一映射的数目将急剧增加, 多对一映射干扰点的去除会把大量正确的匹配点对剔除, 干扰点去除后的候选配准点对中错误点对比例急剧增大, 可致使错误模式成为最终的同名点对, 导致配准结果不准确, 因此文献 [12 13, 16] 并不适合于存在大尺度差异的图像序列配准. 图 3 给出了大尺度变化下采用多对一映射去除干扰点的一个示例, (b) 去除多对一映射的候选同名点对 (b) Corresponding point pairs after removing N -1 mapping (c) RANSAC 算法产生的同名点对 (c) Corresponding point pairs selected by RANSAC 图3 Fig. 3 大尺度变化下采用多对一映射去除的外点去除结果 Removal results of N -1 mapping point pairs 为了保证可见光热红外序列在大尺度变化下也能够精确配准, 本文不采用多对一映射干扰点的

[A,B] t=1 (2) vmin = vm AX low thd vmax = vm AX hgh thd seq (i, j, t) seq (p, q, t) A B t=1 mapa(i, j) = min T P (p,q) U (i,j) V seqa (i, j, t) T P (3) t=1 3) 候选同名点对选取运动状态序列是二值序列, 海明距离是度量这

5 ,,. 2) /,,. ( ).,. 3) RANSAC,, ( Outliers). Fischler Bolles RANSAC [19], 50 %,.,.,,. RANSAC,,. RANSAC 4, Fig. 4 4 RANSAC Our flow chart of cascade RANSAC method 2.3,,.,,., H(, ) x = a 11x + a 12 y + a 13 a 31 x + a 32 y + 1 y = a 21x + a 22 y + a 23 a 31 x + a 32 y + 1 a 11 a 12 a 13 H perspective = a 21 a 22 a 23 (4) a 31 a 32 1, (x, y), (x, y ).,,,. LM [17 18],.,,., LM. E(H), E(H). E(H) : E(H) = N i=1 (x, y ) = H(x, y ) [(x i x i ) 2 + (y i y i ) 2 ] (5), (x i, y i ) i, (x i, y i) i, (x i, y i ) (x i, y i) H(, ), N., E(H) Jacobian,, µ. LM : { H(k + 1) = H(k) + H H = (J T (H)J(H) + µi) 1 J T (H)E(H) E(H) = 1 N e 2 i (H) 2 i=1 (6), J(H) E(H) Jacobian, µ, I. 3,.,

,,., H(, ) x = a 11x + a 12 y + a 13 a 31 x + a 32 y + 1 y = a 21x + a 22 y + a 23 a 31 x + a 32 y + 1 a 11 a 12 a 13 H perspective = a 21 a 22 a 23 (4) a 31 a 32 1, (x,

6 9 : Co-motion 1225,., , 20 /, 400, ,.., (Delay measurement time synchronization, DMTS),. IR913A.,,.. 3.1,, ;, ),, I I II. I ( 5 ). 5 I. 2.1, 6.,, Co-motion. 7,. 50 %.., low thd 0.05, hgh thd (a), ; 8 (b) /,, ; 8 (c) RANSAC, RANSAC,. 10 / 0.3., LM.,. 9 I I (a) (a) Visual image sequence Fig. 5 (b) (b) Thermal image sequence 5 I Image Sequence I in our database

05, hgh thd 0.85. 8 (a), ; 8 (b) /,, ; 8 (c) RANSAC, RANSAC,. 10 / 0.3., LM.,. 9 I I (a) (a) Visual image sequence Fig.

7 (a) (a) Visual motion detection results (b) (b) Thermal motion detection results 6 Fig. 6 Motion detection results (a) (a) Visual accumulated image of motion status vector (b) (b) Visual accumulated image after preselecting (c) (d) (c) Thermal accumulated image of motion (d) Thermal accumulated image after status vector preselecting 7 Fig. 7 Preselecting results of motion status vectors II,, 9 (c). 2) III. 10 III, 9. 3)

accumulated image after preselecting (c) (d) (c) Thermal accumulated image of motion (d) Thermal

8 9期张秀伟等: 基于 Co-motion 的可见光热红外图像序列自动配准算法 (a) 候选同名点对 (b) 基于次优/最优剔除的干扰点 (a) Candidate corresponding points (b) The outlier detected by selection and outlier removal suboptimum/optimum removal 图8 Fig (c) 经级联 RANSAC 算法去除外点后获得的同名点对 (c) Corresponding point pairs selected by cascade RANSAC 候选同名点对选取及外点去除 Candidate corresponding points selection and outlier removal (a) 自建同步序列 I (b) 俄亥俄州大学同步序列 I (a) Mosaic result of an image pair in (b) Mosaic result of Ohio State University Sequence I Sequence I 图9 (c) 自建同步序列 II (c) Mosaic result of Sequence II 平移变化下可见光热红外序列某帧图像的拼接融合结果图 Fig. 9 Registration results under change of translation (a) 可见光图像 (b) 热红外图像 (c) 拼接结果 (a) Visual image (b) Thermal image (c) Mosaics image 图 10 Fig. 10 大尺度变化下的配准结果 (自建同步序列 III) Registration results under the change of large scale (Sequence III) 本文使用自建同步序列 IV 和自建同步序列 V 来说明算法在大尺度旋转变化下的配准效果. 同步序列 IV 为傍晚拍摄, 同步序列 V 为夜间拍摄. 这两组同步序列某一帧的拼接融合结果如图 11 和图 12 所示. 图中的可见光图像和热红外图像间存在着很大的尺度旋转和平移变化, 即便在这种情况下, 拼接融合的结果图中的行人窗户及花坛边缘仍旧得到了很好的对准, 说明了本文算法在大尺度旋转情况下的有效性. 4) 大的视场角变化由于目前公开的红外可见光同步序列数据大多是用于融合, 它们几乎都是同一视角拍摄. 因此, 本文采用自建数据 VI 来说明实验效果. 在拍摄同步序列 VI 时, 热红外相机朝向某建筑的门口拍摄, 而可见光相机斜向该建筑的东南方向拍摄. 图 13 给出了实验结果, 从拼接结果可以看出, 场景中花坛的边沿运动着的行人得到了很好的对准. 在中远场景下本文算法可以适应于较大的视场角变化. 通过平移大尺度较大尺度旋转和大的视场角变化四类实验的分析可知, 本文算法具有很好的鲁棒性, 不易受这四类变化的影响. 需要说明的是, 在大尺度变化下, 配准算法的精确度稍有所下降, 这

(a) Mosaic result of an image pair in (b) Mosaic result of Ohio State University Sequence I Sequence I 图9 (c) 自建同步序列 II (c) Mosaic result of Sequence II 平移变化下可见光热红外序列某帧图像的拼接融合结果图 Fig.

9 ( IV) ((a) ; (b) ; (c) ) Fig. 11 Registration results under the change of large scale and rotation (Sequence IV) ((a) Visual image; (b) Thermal image; (c) Mosaics image) 12 ( V) ((a) ; (b) ; (c) ) Fig. 12 Registration results under the change of large scale and rotation (Sequence V) ((a) Visual image; (b) Thermal image; (c) Mosaics image) 13 ( VI) ((a) ; (b) ; (c) ; (d) ) Fig. 13 Registration results under the change of large viewing angle (Sequence VI) ((a) Visual image; (b) Thermal image; (c) Corresponding points; (d) Mosaics image)., ), 2.1 (3).,.,,. 2).,. III III, III , III III,,. 3),.,,, ( ) ;,,.,

10 9 : Co-motion 1229,,, 2.2,, RANSAC.,.,,.,.,. I, ( 9 (a))., II,,,,. 14 II. 4),. 10,. I,, ; III, ; IV ; VI,,.,,,.,.,,,. I III IV VI,, ( II) ((a) ; (b) ; (c) ; (d) ) Fig. 14 Registration results in the heavy visual shadow situation (Ohio State University Sequence II) ((a) Visual image; (b) Thermal image; (c) Corresponding points; (d) Mosaics image) ,,,.,, ; Ju [11],,,.,,.,.

14 Registration results in the heavy visual shadow situation (Ohio State University Sequence II)

11 ,,,. BaseP ts, H Truth H Alg MatchP ts T MatchP ts A,. 1, 1,, ,. Table 1 1 Accuracy analysis of the proposed method ( x, y) I (0.56, 0.64) II (0.39, 0.44) I (0.74, 1.30) II (0.65, 1.29) III (0.70, 1.85) IV (0.99, 1.92) V (1.34, 1.91) V (1.50, 2.56) (0.86, 1.48),. AMD Athlon64 ( 1.8 GHz), 1 GB, Matlab 7.4.0, 8, , 6.. 4, Co-motion. : 1),,, ; 2) ; 3) ; 4) Co-motion,.,.,. References 1 Krotosky S J, Trivedi M M. Person surveillance using visual and infrared imagery. IEEE Transactions on Circuits and Systems for Video Technology, 2008, 18(8): Chen Y X, Han C Z. Night-time pedestrian detection by visual-infrared video fusion. In: Proceedings of the 7th World Congress on Intelligent Control and Automation. Chongqing, China: IEEE, Zhao J P, Yang H M, Ding Y. Medical image registration algorithm research based on mutual information similarity measure. In: Proceedings of the International Symposium on Photo Electronic Detection and Imaging. Beijing, China: SPIE, Lu X, Zhang S, Su H, Chen Y. Mutual information-based multimodal image registration using a novel joint histogram estimation. Computerized Medical Imaging and Graphics, 2008, 32(3): Gao Feng, Wen Gong-Jian, Lv Jin-Jian. An optimal algorithm for IR/visual image registration based on mainline-pairs. Chinese Journal of Computers, 2007, 30(6): (,,.., 2007, 30(6): ) 6 Hrkac T, Kalafatic Z, Krapac J. Infrared-visual image registration based on corners and Hausdorff distance. In: Proceedings of the 15th Scandinavian Conference. Aalborg, Denmark: Springer, Kim Y S, Lee J H, Ra J B. Multi-sensor image registration based on intensity and edge orientation information. Pattern Recognition, 2008, 41(11): Lei Lin, Jiang Yong-Mei, Kuang Gang-Yao. A method of the remote sensing image registration based on image classification. Journal of National University of Defense Technology, 2004, 26(2): (,,.., 2004, 26(2): 35 40) 9 Caspi Y, Simakov D, Irani M. Feature-based sequence-tosequence matching. International Journal of Computer Vision, 2006, 68(1): Lee L, Romano R, Stein G P. Monitoring activities from multiple video streams: establishing acommon coordinate frame. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2000, 22(8): Ju H, Bhanu B. Fusion of color and infrared video for moving human detection. Pattern Recognition, 2007, 40(6): Szlavik Z, Havasi L, Sziranyi T. Estimation of common groundplane based on co-motion statistics. In: Proceedings of the International Conference on Image Analysis and Recognition. Porto, Portugal: Springer,

Person surveillance using visual and infrared imagery. IEEE Transactions on Circuits and Systems for Video Technology, 2008, 18(8): 1096 1105 2 Chen Y X, Han C Z.

12 9 : Co-motion Szlavik Z, Havasi L, Sziranyi T. Image matching based on co-motion statistics. In: Proceedings of the 2nd International Symposium on 3D Data Processing, Visualization, and Transmission. Thessaloniki, Greece: IEEE, Szlavik Z, Sziranyi T, Havasi L, Benedek C. Optimizing of searching co-motion point-pairs for statistical camera calibration. In: Proceedings of the International Conference on Image Processing. Genoa, Italy: IEEE, Benedek C, Havasi L, Sziranyi T, Szlavik Z. Motion-based flexible camera registration. In: Proceedings of the International Conference on Video and Signal Based Surveillance. Como, Italy: IEEE, Szlavik Z, Sziranyi T, Havasi L. Video camera registration using accumulated co-motion maps. ISPRS Journal of Photogrammetry and Remote Sensing, 2007, 61(5): Levenberg K. A method for the solution of certain nonlinear problems in least squares. Quarterly of Applied Mathematics, 1994, 2(2): Marquardt D W. An algorithm for least-squares estimation of nonlinear parameters. SIAM Journal on Applied Mathematics, 1963, 11(2): Fischler M A, Bolles R C. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981, 24(6): Brisson N M, Zaccarin A. Moving cast shadow detection from a Gaussian mixture shadow model. In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. San Diego, USA: IEEE, Yang Gen, Zhang Yan-Ning, Xi Run-Ping. Detection algorithm for infrared moving objects based on gray delaminating. Infrared and Laser Engineering, 2007, 36(6): (,,.., 2007, 36(6): )... weirui20011@163.com (ZHANG Xiu-Wei Ph. D. candidate at the College of Computer Science, Northwest Polytechnical University. Her research interest covers video content analysis and multi-sensor information cooperative processing. Corresponding author of this paper.).. ynzhang@nwpu.edu.cn (ZHANG Yan-Ning Professor at the College of Computer Science, Northwest Polytechnical University. Her research interest covers digital image processing and pattern recognition.).. yangtaonwpu@163.com (YANG Tao Postdoctor at the College of Computer Science, Northwest Polytechnical University. His research interest covers video content analysis, visual object detection and tracking, and motion video registration and stabilization.).. hermas@163.com (ZHANG Xin-Gong Master student at the College of Computer Science, Northwest Polytechnical University. His research interest covers multisensor information cooperative processing, and motion video registration and stabilization.).. sdp @163.com (SHAO Da-Pei Master student at the College of Computer Science, Northwest Polytechnical University. His research interest covers video content analysis and image recognition.)

Optimizing of searching co-motion point-pairs for statistical camera calibration. In: Proceedings of the International Conference on Image Processing. Genoa, Italy: IEEE, 2005.

,.,, [9] Harris, Hausdorff, ; [10],,,. [11],,,,,., [12 13] ; [14], ;,,.,,., ( ), [15 17],. Tsang [15],,,. Tsang [16] (Simple genetic algorithm

,.,, [9] Harris, Hausdorff, ; [10],,,. [11],,,,,., [12 13] ; [14], ;,,.,,., ( ), [15 17],. Tsang [15],,,. Tsang [16] (Simple genetic algorithm 39 9 Vol. 39, No. 9 2013 9 ACTA AUTOMATICA SINICA September, 2013 1 2 3,,., ;,,, ;,,. LTS Hausdorff (Least trimmed square Hausdorff distance, LTS-HD),.,,.,,. DOI, LTS Hausdorff,,,,.., 2013, 39(9): 1447

9期 张秀伟等: 基于 Co-motion 的可见光 热红外图像序列自动配准算法 (a) 建筑物场景正午时的红外 可见光图像对 (a) Thermal-visual image pair of building scene at noon (b) 道路场景红外 可见光图像对 (b) Thermal-

9期张秀伟等: 基于 Co-motion 的可见光热红外图像序列自动配准算法 (a) 建筑物场景正午时的红外可见光图像对 (a) Thermal-visual image pair of building scene at noon (b) 道路场景红外可见光图像对 (b) Thermal-