-
2006年, da CUNHA等人[10]在contourlet理论[11]的基础上提出了非下采样contourlet变换, 通过图像的多方向、多尺度分解克服了contourlet平移变化导致的频谱重叠等缺陷, 使其具备了很好的特征提取性能。NSCT由非下采样塔式滤波器组(non-subsampled pyramid filter bank, NSPFB)和非下采样方向滤波器组(non-subsampled directional filter bank, NSDFB)构成。NSCT分解是一个迭代过程, 输入图像首先经过NSPFB分解得到与原图大小一致的带通子带、低通子带; 其次NSPFB继续对得到的低通子带进行重复迭代操作。图 1所示为迭代3次的NSPFB分解。
NSCT第二部分由NSDFB将分解得到的所有带通子带在不同方向继续分解, 图 2所示为NSDFB 2级分解。最终输入图像经过NSCT分解后得到一个低通子带和多个子带图像。
NSCT总体分解图如图 3所示。
-
引导滤波(guided image filtering, GIF)[12-13]在图像平滑、增强等领域有较好的效果, 其算法难易程度与滤波窗口半径无关, 可以降低运算量。
本方法局部线性模型为:
$ {q_i} = {a_k}{g_i} + {b_k}, (i \in {\omega _k}) $
(1) 式中, i和k是像素索引, gi是引导图像的值, qi是输出值, ak, bk是当中心窗口ω位于k时该函数的系数。然后由参考文献[14]中引入代价函数, 表示为:
$ E({a_k}, {b_k}) = \sum\limits_{i \in {\omega _k}} {\left[ {{{\left( {{a_k}{g_i} + {b_k} - {p_i}} \right)}^2} + \varepsilon {a_k}^2} \right]} $
(2) 式中, ε是平滑因子, 用来限制ak的取值, pi是输入图像p对应的值。通过最小二乘法, 可以得到ak, bk, 表达式为:
$ {a_k} = \frac{{\frac{1}{{|\omega |}}\sum\limits_{i \in {\omega _k}} {{g_i}{p_i}} - {\mu _k}{{\bar p}_k}}}{{\sigma _k^2 + \varepsilon }} $
(3) $ {b_k} = {{\bar p}_k} - {a_k}{\mu _k} $
(4) 式中, μk和σk2是引导图g的均值和方差, |ω|是窗口包含的像素数, pk是输入图像p在窗口中的均值。如果想具体求某一点的输出值时, 将该点所参与的全部线性函数值平均即可, 表达式如下:
$ {a_i} = \frac{{\sum\limits_{k \in {\omega _k}} {({a_k}{g_i} + {b_k})} }}{{|\omega |}} = {{\bar a}_i}{g_i} + {{\bar b}_i} $
(5) 通过(1)式求得梯度保持函数▽q=a▽g。综上所述, 处理图像q较多的保留引导图g所包含的信息。经过引导滤波处理的图像, 不仅可以平滑背景噪声, 还能有效增强图像边缘, 提高图像信噪比。
-
有学者提出了一种隶属度函数和模糊增强算法[15], 其本质是用某种变换函数将图像映射为一个矩阵, 利用模糊集理论对其进行处理, 但是该算法存在一些缺点, 如渡越点的选取靠多次实验或个人经验, 对实验结果有较大影响, 采用最大类间方差法算法通过自动选取阈值实现渡越点的自动获取[16], 同时为了减少不必要的计算, 简化计算流程, 对算法的模糊隶属函数修改如下:
$ {u_{ij}} = {\{ \sin [\frac{{\pi (f - {f_{\min }})}}{{2({f_{\max }} - {f_{\min }})}}]\} ^r} $
(6) $ {F_1}({\mu _{ij}}) = \left\{ \begin{array}{l} 2{({\mu _{ij}})^2}, (0 \le {\mu _{ij}} \le {u_{\rm{c}}})\\ 1 - 2{(1 - {\mu _{ij}})^2}, ({u_{\rm{c}}} \le {\mu _{ij}} \le 1) \end{array} \right. $
(7) 式中, fmax, fmin对应图像像素中最大值、最小值; r为迭代次数, 可取为1, 2, …; uc是由渡越点决定。然后随着隶属度函数的确认, 通过模糊算子把图像映射到空间域, 最后通过反变换得到处理图像。
-
为了检验本文中的算法在背景抑制方面的效果, 采用天空背景、海天线背景和海面多目标等3组代表性较强的红外图像进行实验, 并用本文中的方法与基于顶帽变换(top-hat)、引导滤波(guided image filter-ing, GIF)、鲁棒性主成分分析(robust principal component analysis, RPCA)和NSCT的4种背景抑制方法比较。使用均方误差(mean-square error, MSE)和峰值信噪比(peak signal-to-noise ratio, PSNR)体现算法在背景抑制方面的性能, MSE用于计算图像背景与预测背景平均误差大小, 误差结果与所计算数值成反比。PSNR所计算数值与目标增强效果成正比。通过对MSE和PSNR的性能指标分析可以得出, 本文中的方法较其它背景抑制算法有一定优势。算法运行环境基于Inter双核3.2GHz, 内存4.0GB的PC机和MATLAB 2014b软件平台。图 5为实验结果。
表 1为5种实验方法对应指标, 本文中的方法与其它4种方法在峰值信噪比方面有了一定的提高, 均方误差显示, 预测背景与真实背景接近。由此表明作者提出的方法对多种复杂红外小目标背景都适用。
Table 1. Comparison among the experimental results
Fig. 5m Fig. 5g Fig. 5a MSE PSNR MSE PSNR MSE PSNR top-hat 16.6446 32.8435 13.0185 36.7370 12.3329 36.1820 GIF 15.5397 33.1418 13.0016 36.7426 12.0448 36.2847 RPCA 10.5324 34.8309 6.6929 36.8000 6.2712 37.0826 NSCT 25.6984 30.9572 10.8570 37.1255 5.8153 39.4470 algorithm of this paper 9.9399 35.0824 5.8548 37.3811 5.1063 40.0116 红外图像在其采集、转存过程都会受到噪声影响, 例如:高斯噪声、斑点噪声、椒盐噪声等一系列噪声, 高斯噪声是一种随机噪声, 其值按高斯概率定律分布; 斑点噪声则是随机散射形成的, 在图像上表现为小斑点, 噪声的存在严重影响图像质量, 因此消除噪声格外重要, 为了验证本文中的算法的鲁棒性, 在天空背景红外图像中分别加入高斯噪声、斑点噪声, 采用本文中的算法对其处理。
图 6为实验结果。图 6a为天空背景图像; 分别加入高斯噪声和斑点噪声, 如图 6b和图 6d所示; 将噪声图像通过本文中的算法处理得到的实验结果如图 6c和图 6e所示, 由于算法在低通子带及带通子带部分分别采用引导滤波和模糊算法, 在平滑噪声的同时加强了目标边缘, 因而能够提取完整地目标区域, 证明了算法的有效性。
基于引导滤波和模糊算法的红外背景抑制算法
Infrared background suppression algorithm based on guided filtering and fuzzy algorithm
-
摘要: 为了减少背景对红外小目标检测结果的影响,同时降低检测虚警率,采用了基于引导滤波和模糊算法的红外背景抑制算法,利用非下采样轮廓波多尺度、多方向的分解机制,将红外序列图像分解为低通子带和带通子带;再利用引导滤波对低通子带处理,以平滑图像、抑制噪声、增强背景细节;带通子带则采用模糊算法处理,实现目标和残留背景分离;最后将各子带图像通过非下采样轮廓波逆变换,得到了背景抑制图像。结果表明,该方法可以将均方误差降至5~10,有效抑制了背景,突出了目标。该研究为提高复杂背景下的红外小目标检测精度提供了支持。Abstract: In order to reduce the influence of background on detection results of infrared small targets and reduce false alarm rate, infrared background suppression algorithm based on guidance filter and fuzzy algorithm was adopted. The infrared image was decomposed into low-pass band and band-pass band by using multi-scale and multi-direction decomposition mechanism of non sampled contour. The guided filter was used to process low-pass sub-band to smooth images, suppress noise and enhance background details. Band-pass sub-band was processed by fuzzy algorithm to seperate the target from the residual background. Background suppression image was obtained by changing subband images through non subsampled contour inversion. The results show that, the method can reduce mean square error to 5~10, and effectively suppress the background and highlight the target. This study provides the support for improving the detection accuracy of infrared small targets in complex background.
-
Table 1. Comparison among the experimental results
Fig. 5m Fig. 5g Fig. 5a MSE PSNR MSE PSNR MSE PSNR top-hat 16.6446 32.8435 13.0185 36.7370 12.3329 36.1820 GIF 15.5397 33.1418 13.0016 36.7426 12.0448 36.2847 RPCA 10.5324 34.8309 6.6929 36.8000 6.2712 37.0826 NSCT 25.6984 30.9572 10.8570 37.1255 5.8153 39.4470 algorithm of this paper 9.9399 35.0824 5.8548 37.3811 5.1063 40.0116 -
[1] RONG Sh H, LIU G, ZHOU H X, et al. Infrared dim and small target background suppression based on the improved shearlet transformand the guide filter[J]. Acta Photonica Sinica, 2015, 44(2):210002(in Chinese). doi: 10.3788/gzxb [2] YAN G Sh, BI W Zh. Detection algorithm of small target based on regional singularity filter[J]. Optical Technique, 2007, 33(2):163-152(in Chinese). [3] QIN H L, ZHOU H X, LIU Sh Q, et al. Total variation for dim and small target background suppression[J]. Optical Technique, 2009, 35(4):596-598(in Chinese). [4] QIN H L, YAO K K, ZHOU H X, et al. Nonsubsampled directional filter banks for infrared complex ground background suppression[J]. Semiconductor Optoelectronics, 2011, 32(4):560-563(in Ch-inese). [5] LI J, MA J N, LI Sh J, et al. Background suppression for infrared dim small target detection based on spatial-temporal filter[J]. Semiconductor Optoelectronics, 2017, 38(3):396-400(in Chinese). [6] GE W, JI P Ch, ZHAO T Ch. Infrared and visible light images fusion of fuzzy logic on NSST domain[J]. Laser Technology, 2016, 40(6):892-896(in Chinese). [7] WU Y Q, LUO Z J, WU W Y. A method of small target detection in infrared image based on non-subsampled contourlet transform[J]. Journal of Image & Graphics, 2009, 14(3):477-481(in Chinese). [8] WU T A, HUANG Sh C, YUAN Zh W, et al. NSCT combined with SVD for infrared dim target complex background suppression[J]. Infrared Technology, 2016, 38(9):758-764(in Chinese). [9] FENG Y. Detection of dim and small infrared targets based on the improved singular value decomposition[J]. Laser Technology, 2016, 40(3):335-338(in Chinese). [10] da CUNHA A L, ZHOU J, DO M N. The non-subsampled contourlet transform:theory, design and applications[J]. IEEE Transactions on Image Proecessing, 2006, 15(10):3089-3101. doi: 10.1109/TIP.2006.877507 [11] PENG Zh, ZHAO B J. Novel scheme for infrared image enhancement based on contourlet transform and fuzzy theory[J]. Laser & Infrared, 2011, 41(6):635-640(in Chinese). [12] HE K, SUN J, TANG X OU. Guided image filtering[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2013, 35(6):1397-1409. [13] LI X H, DONG A G, FENG J H. Regional fusion algorithm of i-mages based on multistage guide filters[J]. Laser Technology, 2016, 40(5):756-761(in Chinese). [14] KANG Ch Q, CAO W P, HUA L, et al. Infrared image denoising algorithm via two-stage 3-D filtering[J]. Laser & Infrared, 2013, 43(3):261-264(in Chinese). [15] ZHANG W. An improved method of image fuzzy enhancement[J]. Life Science Instruments, 2006, 4(3):29-31(in Chinese). [16] LUO Y, YANG H M. An improved fuzzy enhancement algorithm of robot digital landmark based on Otsu adaptive threshold[J]. Optics & Optoelectronic Technology, 2010, 8(1):35-38(in Chinese).