-
选用参考文献[15]中的中波红外图像数据(MWIR)对所提方法进行测试。图2给出了该数据集中包含的10类目标的典例图像,既有民用轿车也有军用的坦克、装甲车等。不同目标在工作状态下的红外热效应可从图中直观看出。对于任一类目标,MWIR数据集通过圆周观测的方式获得不同方位角的图像结果。为开展分类实验,选取每一类目标的120幅图像作为训练样本,100幅图像作为测试样本。
为验证所提方法的有效性,将其在相同测试条件下与现有红外识别方法进行对比。4类对比方法的相关信息如表1所示。这4类方法既包括特征主导类别(保留轮廓特征、HOG特征、协方差描述子)也包括分类器主导类别(SVM、核稀疏编码),还包含基于深度学习模型(CNN)的新方法,具有较强的对比意义,可有效反映所提方法的性能。后续测试除在原始样本上开展外,还模拟构建了噪声场景和遮挡场景,进一步评估所提方法的稳健性。
-
实验首先基于原始样本开展,按照文中设置10类目标的测试样本各有100幅。所提方法对各类目标的识别结果展示如图3所示,描述了测试样本被归属为不同类别的样本个数。其中,对角线元素即为正确识别的样本数。定义10类目标的平均识别率为所有正确识别样本占全体测试样本的比例,计算可得所提方法的结果为98.2%,处于较高的水平。表2在相同条件下比较了所提方法和4类对比方法的性能。从分类精度(即平均识别率)来看,所提方法性能最佳。基于CNN的对比方法4仅次于所提方法,表明深度学习模型很好的分类能力。其余3类对比方法尽管可以取得94%以上的平均识别率,但相对所提方法仍有一定差距。从分类效率来看,表2给出的不同方法进行单个测试样本的处理时间表明,所提方法相比直接运用稀疏表示或SVM的两类对比方法效率略低,但整体处于较高的水平。综合考虑分类精度和效率,所提方法对于原始样本的分类能力较强,整体优势较为明显。
Method type Average recognition rate Efficiency/ms Proposed method 98.2 94.8 Reference 1 94.4 146.7 Reference 2 95.1 87.3 Reference 3 95.3 92.4 Reference 4 97.4 113.2 Table 2. Comparison of performance of different methods on orginal test samples
-
图2所示的原始样本获取条件较为理想,目标的完整度和图像的信噪比(Signal-to-noise ratio,SNR)都较高。实际应用场景下,由于可能受到自然或人为干扰,红外图像的信噪比往往较低。该实验在原始样本的基础上,通过添加噪声的形式构造不同信噪比的测试集,验证所提方法的噪声稳健性。具体地,以原始样本能量为参照,根据预设的信噪比生成加性高斯白噪声,进而获得相应的噪声样本。按照上述思路,文中分别构造信噪比为10 dB、5 dB、0 dB、−5 dB和−10 dB的测试集。对各类方法的测试结果统计如图4所示。直观可见,噪声干扰对红外目标分类结果产生了显著的影响。对比而言,所提方法可在不同信噪比保持最高的平均识别率,体现其稳健性。在4类对比方法中,基于轮廓匹配的对比方法1整体稳健性最强,稀疏编码方法次之,体现了几何形状特征以及稀疏分类机制对于噪声干扰的一定稳健性。所提方法在特征提取阶段运用单演信号获得多层次互补的特征矢量,可有效克服噪声影响;在分类阶段运用基于稀疏表示原理的联合稀疏表示模型,进一步增强了结果的噪声稳健性。
-
实际获得的目标红外图像还可能出现遮挡的情形。此时,可用于特征提取和分类的信息完整度受到影响。该实验主要考察所提方法在不同遮挡条件下的性能。具体地,以目标被遮挡的比例作为基础参数,模拟构造10%、20%、30%、40%和50%遮挡条件下的4组测试集。对各类方法在遮挡条件下的分类性能进行测试,获得如图5所示的结果。相比噪声干扰的情形,遮挡对于红外目标分类性能的整体影响更为剧烈。所提方法在各个遮挡比例下均保持最高性能,体现其稳健性。与噪声干扰相似,基于轮廓匹配和稀疏编码的两类方法在对比方法中处于较优的水平,反映特征及分类器的有效性。所提方法通过结合单演信号和联合稀疏表示模型的优点提高了遮挡条件下的整体分类性能。
Research on monogenic signal of application in infrared imagery target classification
doi: 10.3788/IRLA20210165
- Received Date: 2021-05-10
- Rev Recd Date: 2021-07-20
- Publish Date: 2021-12-31
-
Key words:
- infrared imaging /
- target classification /
- monogenic signal /
- joint sparse representation
Abstract: Infrared imaging is an important measure for night observation, which is widely used in both military and civil fields. For infrared imagery target classification problem, the monogenic signal was introduced for feature extraction, which was used to analyze the target characteristics. The infrared image after single signal processing can be described by the amplitude, phase and orientation components. For each component, its multi-scale decompositions were processed by connection and downsampling to achieve one single feature vector. Finally, three feature vectors were generated to describe the multi-layer properties of the target. The joint sparse representation was employed as the representation model for the three feature vectors, which used their correlation to improve the overall reconstruction precision. The reconstruction errors of different classes were calculated based on the results from joint sparse representation and the target label could be further decided. The experiments were conducted on the medium wave infrared (MWIR) image dataset to classify the original, noisy, and occluded samples. By comparison with several existed algorithms, the validity and robustness of the proposed method for infrared imagery target classification could be confirmed.