-
光谱聚类是将具有一定程度相同属性的光谱信息分为一类,按照类似规则进行分类,形成不同类别组的过程[11]。所以建立不同的相似度函数可以使数据分类更加准确。对于该系统而言就是要得到最佳的光谱分类,从而提高目标在复杂环境中的信噪比[12]。由于实际目标往往存在固定的几何外形,但目标的边界有可能被遮挡或与背景的辐射相近而造成无法区分,故文中算法融合了目标空间信息和目标光谱信息,将几何特征作为边界条件限制光谱聚类过程,从而实现目标区域的去伪滤波,提高系统对传统红外图像中目标的识别能力。
对于每个可分辨的光谱区间而言,任意两个像素点间的相似度函数[13]都利用谱聚类的形式进行分类。设像素点分别为xi与xj,由此可以推导得到高斯径向基函数[14]G为:
$$G = \exp \left( { - \frac{{{{\left\| {{p_i} - {p_j}} \right\|}^2}}}{{\sigma _p^2}}} \right)$$ (1) 式中:σP为参数标量;pi和pj为像素点xi与xj中对应的灰度值,i和j分别为像素点对应的序号。
由于系统需要将几何数据中的特征信息与光谱数据中的特征信息相融合,是不同物理量纲之间的数据的制约关系,故采用不受量纲影响的马氏距离可以有效解决由于数据属性不同造成的数据不匹配,并且采用马氏距离对数据分析还能排除变量间互扰的问题。故系统对应的马氏距离dm为:
$${d_m} = \exp \left( { - {{\left( {{s_i} - {s_j}} \right)}^{\rm T}}{\boldsymbol{A}}\left( {{s_i} - {s_j}} \right)} \right)$$ (2) 式中:A表示全矩阵;si与sj为像素点i和j中的光谱特征值。
由此将公式(1)、(2)相乘获得相似度函数:
$$d\left( {{x_i},{x_j}} \right) = G \cdot {d_m}$$ (3) 由此可见,通过公式(3)可以对两个不同光谱段的两个空间点构造相似度函数,从而实现通过空间关系与光谱特性关系的方式提高目标的识别能力。
Infrared target recognition algorithm based on bounding box constrained spectral clustering
-
摘要: 在红外成像过程中,目标边缘模糊化是影响红外目标识别效果的关键因素,也是红外目标识别算法的研究重点,故在光谱图像中合理补偿目标几何特征信息成为研究热点之一。结合包含目标几何特征信息的包围盒作为约束条件,对红外光谱图像进行分层限定滤波,降低原有图像数据中目标几何外形数据的丢失,提高目标可识别性。设计了在包围盒约束条件下的光谱聚类算法,设置参数η表征待测军用车辆目标的几何信息,设置参数m表征待测军用车辆目标的光谱特征信息。实验采用TEL-1000-MW型红外成像光谱仪获取多光谱图像,通过改变m和η值调整光谱特征值个数与包围盒范围,从而获得不同的目标识别图像。并与传统方法对同一幅红外目标图像的识别效果相比较,结果发现采用包围盒约束的待测目标图像几何边界信息保留效果明显优于传统方法,当m=10、η=0.7时,红外图像的目标识别效果最好,同时算法收敛速度也最优。由此可见,该算法在提高红外目标识别能力、避免误判伪目标和漏检目标方面具有很高的实用价值。Abstract: In the process of infrared imaging, target edge blurring is a key factor that affects the effect of infrared target recognition, and it is also the focus of infrared target recognition algorithms. Therefore, reasonable compensation of target geometric feature information in spectral images has become one of the research hotspots.The bounding box containing the geometric feature information of the target was used as a constraint condition, and the infrared spectrum image was hierarchically limited and filtered to reduce the loss of the target geometric shape data in the original image data and improve the recognizability of the target. A spectral clustering algorithm under bounding box constraints was designed. The parameter η was set to characterize the geometric information of the military vehicle target under test, and the parameter m was set to characterize the spectral feature information of the military vehicle target under test. In the experiment, a TEL-1000-MW infrared imaging spectrometer was used to obtain multi-spectral images. By changing the m and η values, the number of spectral feature values and the bounding box range were adjusted to obtain different target recognition images. Compared with the traditional method for the recognition effect of the same infrared target image, it was found that the geometric boundary information retention effect of the target image under test using the bounding box constraint was significantly better than that of the traditional method. When m=10, η=0.7, the infrared image target recognition effect was the best, and the algorithm convergence speed was also the best. It can be seen that the algorithm has high practical value in improving the ability of infrared target recognition and avoiding false targets and missed targets.
-
-
[1] Shu R, Zhou Y P, Lu C L. Best detection wavelength bands selection method based on multispectral radiation difference [J]. Infrared and Laser Engineering, 2014, 43(8): 2505-2512. (in Chinese) [2] Zhang P, Wang L, Huang W, et al. Multiple pedestrian localization based on couple-states Markov chain with semantic topic learning for video surveillance [J]. Soft Computing, 2015, 19(1): 85-97. (in Chinese) doi: 10.1007/s00500-014-1375-9 [3] Huang W, Yang W J, Zeng J, et al. A novel algorithm of moving object detection via spectral clustering and incremental learning [J]. Journal of Northwestern Polytechnical University, 2017, 35(1): 170-175. (in Chinese) [4] Wang C L, Wang H W, Hu B L, et al. A new spectral-spatial algorithm method for hyperspectral image target detection [J]. Spectroscopy and Spectral Analysis, 2016, 36(4): 1163-1169. (in Chinese) [5] Yu C Y, Xie J L, Fei B, et al. The effect of ambient temperature on human head’s surface skin temperature [J]. Spectroscopy and Spectral Analysis, 2017, 37(1): 172-176. (in Chinese) [6] Sun Y F, Chang X G, Li D X, et al. Infrared image edge detection algorithm based on adaptive Canny [J]. Journal of Shandong University of Technology (Natural Science Edition), 2017, 31(6): 18-21. (in Chinese) [7] Lang Y, Yuan B. Algorithm application based on the infrared image in unmanned ship target image recognition [J]. Microprocessors and Microsystems, 2021, 80: 103554. doi: 10.1016/j.micpro.2020.103554 [8] Singh S, Rao D V. Recognition and identification of target images using feature based retrieval in UAV missions [C]//Computer Vision, Pattern Recognition, Image Processing & Graphics, IEEE, 2014. [9] Xiao Y S, Huang L Z, Zhou J J. RATR of adaptive angular-sector segmentation based on grey incidence analysis model [J]. Grey Systems: Theory and Application, 2017, 7(1): 71-79. doi: 10.1108/GS-09-2016-0034 [10] Cao W, Zhou H, Zhou Z M, et al. An approach for high resolution radar target recognition based on BP neural network [C]//International Conference on Intelligent Computing, ICIC 2011: Advanced Intelligent Computing, 2011: 33-39. [11] Zhou D Y. Radar target HRRP recognition based on reconstructive and discriminative dictionary learning [J]. Signal Processing, 2016, 126(11): 52-64. [12] Huang X Y, Nie X L, Hong W W, et al. SAR target configuration recognition based on the biologically inspired model [J]. Neurocomputing, 2017, 234(4): 185-191. [13] Yuan P, Mao J L, Xiang F H, et al. Improved network fault diagnosis based on genetic optimization BP neural network [J]. Power System and Automation Journal, 2017, 29(1): 118-122. (in Chinese) [14] Wu Zh J, Niu M, Xu B, et al. Research on recognition method based on spectral regression feature reduction and backward propagation neural network [J]. Journal of Electronic and Information, 2016, 38(4): 978-984. (in Chinese) [15] Xu B, Chen B H, Liu H W, et al. Based on the recurrent neural network model, radar high resolution distance image target recognition [J]. Journal of Electronics and Information, 2016, 38(12): 2988-2995. (in Chinese)