-
图1所示为经典LCM算法的过程,A为红外图像,
${{{B}}^k}$ 为滑动块,均分为9块,其中第0块为目标块,1~8为背景块。滑动操作从右至左、从上至下依次遍历图像。计算背景块的均值:$$ {\hat b_l}^{k - {\text{LCM}}}{\text{ = }}\sum\limits_{i = 1}^M {\sum\limits_{j = 1}^M {A_{i,j}^k} } /M,l = 1,\cdots,8 $$ (1) 定义经典LCM算法的异常检测指标:
$$ C_{{\text{LCM}}}^k = \mathop {\min }\limits_l \frac{{{{\left( {{L^k}} \right)}^2}}}{{\hat b_l^{k - {\text{LCM}}}}} $$ (2) 式中:
$ {L^k} $ 为第k滑动块中目标块(即第0模块)的强度最大值。基于经典LCM算法的特点可知,该方法通过对比度值的加权提高信噪比和检测概率。但依然存在两个缺点:(1)虚警点会增加,这是由于取极大值的局限性导致,当噪声为孤立点且值较大时,由于对比度较大,使用滑动窗口计算时会得到一片局部虚警区域;(2)鲁棒性较差,当噪声较强时,由于目标块中采用的是取极大值方式计算,因此,经典LCM算法极易受噪声干扰,文献提出一种基于空间距离改进的视觉显著性弱小目标检测算法(简称VSI-SD LCM),可以有效改善经典LCM算法的以上两个缺点。这种方法考虑视觉注意机制中的空间关系,根据欧几里德距离加权处理,给远距离区域分配较小的权值,得到了较好的效果。但是,这种改良的方法是一种线性滤波方法,在抑制孤立噪声点时效果依然还有待进一步提高。
-
人眼进行异常检测的机制为:视窗内孤立噪声点不影响人类对整体区域亮度的判定,该机制对于背景中有效分离弱小目标特别有启发意义。综合考虑VSI-SD LCM算法的一些缺点,并参考人眼生物视觉方面的一些优势,提出一种基于非线性抗干扰滤波的视觉显著性弱小目标检测算法(简称NLA LCM),图2为算法流程图。
图2中,在VSI-SD LCM算法的基础上,设置滑动窗口在图像中进行移动窗口滤波。窗口将图像ROI内容分为两部分—目标模块及背景模块。定义提出算法的异常检测指标:
$$ C_{{\text{NLA LCM}}}^k = \frac{{{\hat b_0}^k}}{{\hat b_{other}^k}} $$ (3) 式中:
${ \hat b}_{other}^k$ 为背景块的非线性估计值;${\hat b}_0^k$ 为第k个滑动块中目标块的非线性估计值(见图2),公式为:$$ \hat b_0^k{\text{ = }}{\varTheta _{\text{0}}}\left( {A_{i,j}^k} \right) $$ (4) 式中:
$\varTheta$ 为求中值运算符;$ {A_{i,j}^k} $ 为图像在坐标i,j上的强度值,图像块为第k点的第0目标图像块,采用排序的方式排列像素强度值,并取中值。基于目标块和背景块估计值定义的异常检测指标,可有效削弱噪声影响。背景块的非线性估计值的计算式为:
$$ {\hat b}_{other}^k{\text{ = }}\varTheta \left( {q_l^k \cdot {\hat b}_l^k} \right),l = 1,\cdots,8 5 $$ (5) 式中:l表示剩余的8个背景块的序号;
$ q_l^k $ 为第l块的权值,与距离中心像素块的欧几里德距离相关,成反比;${\hat b}_l^k$ 为第l背景块的强度非线性估计值,表达式如下:$$ {\hat b}_l^k{\text{ = }}{\varTheta _l}\left( {A_{i,j}^k} \right) $$ (6) 构建基于中值的公式计算的原因是,有效避免噪声点对不同的背景像素块的影响,消除背景像素块的估计偏差,降低噪声对视觉显著性的影响。但基于均值加权的方法并不能完全消除噪声的影响,仅仅只能降低对异常指标检测的影响。而基于中值排序的方法是一种非线性的方法,在对抗噪声方面优势很大。
使用距离的优势是基于HVS考虑,对比度的表现不仅与强度值相关,而且与距离相关,且距离越远,相关性越小。计算公式如下:
$$ q_l^k{{ = }}q_l^k{\rm{ = }}\dfrac{{1/d_l^k}}{{\displaystyle\sum\limits_{k = 1}^8 {1/d_l^k} }} $$ (7) 式中:
$ d_l^k $ 为第k个背景像素块中心像素点与第0块目标像素块中心之间的欧几里德距离。最后,对显著性图进行自适应阈值分割,可以分割出目标,公式如下:
$$ T{\text{ = }}E{\text{ + }}\lambda D 8 $$ (8) 式中:E和D分别为局部区域的图像灰度期望值和标准差值;
$ \lambda $ 为人工设置的权值系数。实验表明窗口与目标尺度必须匹配,为目标大小的3倍左右较宜。 -
实验采用512×480 pixel的图像进行测试。算法采用Matlab软件编写,配置为3 GHz Intel i5处理器双核电脑完成处理,背景图像选用天空背景、海洋背景、地面背景以及海天背景。定义信噪比SNR为局部区域目标的灰度均值比背景的灰度均值。使用文中提出的NLA LCM算法、经典LCM算法和VSI-SD LCM[17]算法进行比较。
考虑红外成像系统的点扩散效应,理想目标点成像后会模糊,呈高斯分布。目标点的大小约为3×3 pixel,如图3所示。文中对三种算法均选取滑动窗口的大小为15×15 pixel,即目标块窗口大小为5×5 pixel。
在目标检测效果上,NLA LCM算法和VSI-SD LCM算法检测效果较佳,而经典LCM算法会增加目标检测的虚警点。
进一步,使用信噪比及背景噪声抑制因子定量描述三种算法的检测结果[18-19],定义如下:
$$ {{SNR = }}\frac{{\overline {{{{I}}_{{S}}}} }}{{{\sigma _{{N}}}}} $$ (9) $$ {{BF = }}\frac{{{\sigma _{{B}}}}}{{{\sigma _{{A}}}}} $$ (10) 式中:SNR为信噪比;
$\overline {{{{I}}_{{S}}}}$ 为目标信号的均值;${\sigma _{{N}}}$ 为背景噪声的标准差;BF为背景抑制因子;${\sigma _{{B}}}$ 为背景噪声在算法处理前的标准差;${\sigma _{{N}}}$ 为背景噪声在算法处理后的标准差。图4为在红外海背景图像中进行目标检测的效果对比图。其中图(a)为原始图像,目标在图3中已标注出。如果不标注出目标位置,不经提示,人眼很难明显分辨出目标与海背景中的杂波干扰。图(b1)~(b3)分别为传统LCM算法、VSI-SD LCM算法以及文中提出的NLA LCM算法参与检测的显著图。可见前两种算法对背景杂波及干扰噪声的抑制能力有限。目标依然很难从背景中分离出来,而文中提出算法在显著图中明显效果强于其他两种算法,能够有效抑制背景噪声。图(c1)~(c3)分别为传统LCM算法、VSI-SD LCM算法以及文中提出的NLA LCM算法显著图进行二值化后的图像。很明显,最易检测目标的图像为图(c3),目标与杂波及干扰被有效区分,检测目标难度最小,验证了提出算法的有效性。
表1所示为4组图像的信噪比及背景噪声抑制因子使用三种算法处理后的指标对比,通过对比,文中提出的NLA LCM算法处理后信噪比得到了较大改善,背景抑制能力也最强。
表 1 三种算法的指标对比
Table 1. Comparisons of three algorithms
No. Original image LCM VSI-SD LCM NLA LCM Image sequences 1 3.2/1.0 4.5/2.6 7.3/5.9 8.1/6.0 Image sequences 2 2.1/1.0 6.5/4.3 8.3/5.2 10.6/9.3 Image sequences 3 5.0/1.0 11.8/8.4 15.0/13.9 29.5/26.7 Image sequences 4 7.9/1.0 24.1/19.6 35.1/28.4 47.3/35.8 注:表格中数据表示SNR/BF 为进一步验证算法性能,使用ROC曲线进行进一步评价。该曲线以虚警率和检测率作为横纵坐标绘制,描述了目标检测性能。
使用真实的红外图像进行评价结果如图5(a)所示,加噪声后结果如图5(b)所示。可见,在真实红外图像中,检测性能以文中提出的NLA LCM算法为最佳,具备较高的检测率和较低的虚警率。加噪声后,NLA LCM算法性能下降较慢,但VSI-SD LCM算法及经典LCM算法性能下降严重,验证了文中提出算法的抗噪声能力较强。
针对提出算法的处理速度进行对比,依然在四组图像中测试处理速度,如表2所示。可见,提出算法的速度明显比其他两种算法慢,需在软硬件上进行大量优化,提升其实时性能,才能满足需求。
表 2 三种算法的处理速度对比
Table 2. Comparison of processing speed of three algorithms
No. LCM VSI-SD LCM NLA LCM Image sequences 1 0.839 0.592 1.474 Image sequences 2 0.114 0.367 1.598 Image sequences 3 0.139 0.770 1.472 Image sequences 4 0.252 0.270 1.176
Visual salient dim small target detection based on nonlinear anti noise estimation
-
摘要: 针对红外图像处理技术中弱小目标检测的重要性及关键性,提出一种基于非线性抗噪声估计的检测算法来解决高可靠性、高鲁棒性的弱小目标检测问题。提出的方法基于传统视觉显著度算法及空间距离处理方法,对目标及背景区域采用非线性加权方法进行估计,在不显著降低目标信号信噪比的基础上,削弱孤立微小噪声点对检测算法性能的影响,可提高抗噪性能。首先,采用模块化及非线性映射方式预测背景;然后,融入距离相关因子滤除噪声干扰;最后,在处理结束的图像上进行二值化阈值分割,自动检测并向下一级处理软件输出目标位置信息。实验结果表明:提出的算法与近年来先进的弱小目标检测算法相比,在受试者测试曲线上,在相同的虚警率下,可获得更高的检测率,对背景噪声的抑制很明显;在局部信噪比及背景抑制因子的测试比对数据上,提出的算法可获得更高的检测指标。缺点是算法采用了非线性处理技术,运算效率较低,需进一步优化算法以提高计算速度,实现算法的实时目标检测。Abstract: Aiming at the importance and key of dim small target detection in infrared image processing technology, a detection algorithm based on nonlinear anti noise estimation is proposed to solve the problem of dim small target detection with high reliability and robustness. Based on the traditional visual saliency algorithm and spatial distance processing method, the proposed method uses the nonlinear weighting method to estimate the target and background area. On the basis of not significantly reducing the signal-to-noise ratio of the target signal, the influence of isolated small noise points on the performance of the detection algorithm can be weakened, and the anti-noise performance can be improved. Firstly, the background is predicted by modular and nonlinear mapping, and then the distance correlation factor is integrated to filter out the noise interference. Finally, the binary threshold segmentation is carried out on the processed image to automatically detect and output the target position information to the next level processing software. The experimental results show that the proposed algorithm can obtain a higher detection rate on the subject test curve under the same false alarm rate and significantly suppress the background noise compared with the advanced weak and small target detection algorithm in recent years; On the test comparison data of local signal-to-noise ratio and background suppression factor, the proposed algorithm can obtain higher detection indexes. The disadvantage is that the algorithm adopts nonlinear processing technology and has low operation efficiency. It needs to further optimize the algorithm to improve the calculation speed and realize the real-time target detection of the algorithm.
-
表 1 三种算法的指标对比
Table 1. Comparisons of three algorithms
No. Original image LCM VSI-SD LCM NLA LCM Image sequences 1 3.2/1.0 4.5/2.6 7.3/5.9 8.1/6.0 Image sequences 2 2.1/1.0 6.5/4.3 8.3/5.2 10.6/9.3 Image sequences 3 5.0/1.0 11.8/8.4 15.0/13.9 29.5/26.7 Image sequences 4 7.9/1.0 24.1/19.6 35.1/28.4 47.3/35.8 注:表格中数据表示SNR/BF 表 2 三种算法的处理速度对比
Table 2. Comparison of processing speed of three algorithms
No. LCM VSI-SD LCM NLA LCM Image sequences 1 0.839 0.592 1.474 Image sequences 2 0.114 0.367 1.598 Image sequences 3 0.139 0.770 1.472 Image sequences 4 0.252 0.270 1.176 -
[1] Fan M, Tian S, Liu K, et al. Infrared small target detection based on region proposal and CNN classifer [J]. Signal, Image and Video Processing, 2021, 182: 961-973. [2] Ju M, Luo J, Liu G, et al. A real-time small target detection network [J]. Signal, Image and Video Processing, 2021, 15: 1265-1273. [3] Peng Chaoliang, Chu Xianwu, Wang Yunming, et al. Small target detection algorithm in remote sensing image based on improved Yolo[C]//IWAMA, 2021: 612-620. [4] Zhang Xiaoning, Wang Yuqing, Chen Xiaolu. Multi-path recurrent enhanced salient object detection method [J]. Foreign Electronic Measurement Technology, 2021, 40(5): 1-7. (in Chinese) [5] Zhong Lei, Li Yong, Mou Zhiying, et al. Detection method for weak target under unknown strong clutter based on DP-TBD [J]. Systems Engineering and Electronics, 2019, 41(1): 43-49. (in Chinese) [6] Li Dexin, Zhong Hong. Infrared dim small target detection algorithm based on difference histogram of oriented gradients and saliency map [J]. Optical Technique, 2021, 47(5): 594-600. (in Chinese) [7] Wei Shuigen, Wang Chengwei, Chen Zhen, et al. Infrared dim target detection based on human visual mechanism [J]. Acta Photonica Sinica, 2021, 50(1): 0110001. (in Chinese) [8] Cai Wei, Xu Peiwei, Yang Zhiyong, et al. Dim-small targets detection of infrared images in complex backgrounds [J]. Journal of Applied Optics, 2021, 42(4): 643-650. (in Chinese) doi: 10.5768/JAO202142.0402002 [9] Zhang Xiangyue, Ding Qinghai, Luo Haibo, et al. Infrared dim target detection algorithm based on improved LCM [J]. Infrared and Laser Engineering, 2017, 46(7): 0726001. (in Chinese) doi: 10.3788/IRLA201746.0726001 [10] Liu Fen, Cheng Yongce, Zheng Yao, et al. Research on the algorithm of infrared dim small target detection based on SCNN [J]. Laser & Infrared, 2021, 51(6): 820-824. (in Chinese) doi: 10.3969/j.issn.1001-5078.2021.06.021 [11] Tian Wen, Li Suruo. Infrared dim target detection algorithm based on frequency-space domain mapping and multi-scale top-hat transform [J]. Optical Technology, 2018, 44(3): 325-332. (in Chinese) [12] Zhou Yuan, Zhang Jianmin, Lin Xiao. Infrared small target detection using weighting LoG operator [J]. Journal of Applied Optics, 2017, 38(1): 114-119. (in Chinese) [13] Yang Qili, Zhou Binghong, Zheng Wei, et al. Small infrared target detection based on fully convolutional network [J]. Infrared Technology, 2021, 43(4): 349-356. (in Chinese) [14] Ma Yuying, Huang Chengzhang, Huang Yingying, et al. Research of infrared dim and small targets detection technology based on SSD framework [J]. Laser & Infrared, 2021, 51(10): 1342-1347. (in Chinese) doi: 10.3969/j.issn.1001-5078.2021.10.013 [15] Yu Moran, Luo Haibo, Liu Guangqi, et al. Infrared dim and small target detection network based on spatial attention mechanism [J]. Optics and Precision Engineering, 2021, 29(4): 843-853. (in Chinese) doi: 10.37188/OPE.20212904.0843 [16] Deng Jianxun, Xiong Zhongyang, Deng Xin. Dim target detection in air based on information fusion [J]. Electronics Optics & Control, 2018, 25(2): 5-10. (in Chinese) doi: 10.3969/j.issn.1671-637X.2018.02.001 [17] Yang Linna, An Wei, Lin Zaipin, et al. Small target detection based on visual saliency improved by spatial distance [J]. Acta Optica Sinica, 2015, 35(7): 0715004. (in Chinese) doi: 10.3788/AOS201535.0715001 [18] Zhai Hao, Luo Xiaolin, Wu Linxia, et al. Infrared dim target detection based on improved weighted kernel norm [J]. Laser & Infrared, 2021, 51(6): 776-781. (in Chinese) doi: 10.3969/j.issn.1001-5078.2021.06.014 [19] Han Jinhui, Dong Xinghao, Jiang Yawei, et al. Infrared small dim target detection based on local contrast mechanism [J]. Infrared Technology, 2021, 43(4): 357-366. (in Chinese)