-
文中利用红外热像仪采集的车辆目标的红外图像作为基础样本,通过预处理获取目标区域切片,构建训练和测试集。图2所示为实验中涉及的五类目标,包括三轮车、摩托车、小货车、卡车和轿车。所有图像样本经适当裁剪后具备同一尺寸。对于各类目标,选用80幅不同条件下获取的图像作为训练样本,50幅目标图像作为测试样本。
实验过程中,为充分验证提出方法的性能,选用现有几类红外图像目标识别方法进行对比分析。具体对比算法包括参考文献[11]中基于SRC的方法;参考文献[14]中基于CNN的方法;参考文献[16]联合深度特征的方法(记为JSRDeep)以及参考文献[5]中采用局部纹理特征的方法(记为Texture)。其中,SRC和CNN方法仅仅为文中方法的一部分,可通过结果对比直观反映文中分类器决策融合的实际效能。
-
首先在原始训练和测试样本的基础上对方法性能进行初步验证。如图2所示,原始样本的获取条件相对良好,图像中目标均为完整,且信噪比(signal-to-noise ratio,SNR)较高,因此识别问题的难度相对较小。表1所示为文中方法对于五类目标测试样本的详细识别结果,其中Tar1、Tar2、Tar3、Tar4和Tar5依次对应图2中的五类目标。根据表1计算得出它们对应的正确识别率分别为96%、100%、96%、98%、98%,平均识别率为97.6%。这一结果反映了提出方法对于红外目标识别的有效性。在相同的场景下,对4类对比方法进行了相同实验,得到SRC、CNN、联合深度以及局部纹理特征的平均识别率分别为95.8%、96.9%、97.2%、96.3%。对比可见,文中方法的识别性能优于几类对比方法,显示其性能优势。特别地,与SRC和CNN两类方法相比,文中正是对它们进行了联合运用和层次化的决策融合,进一步提升了识别性能,表明了文中决策融合算法的有效性。
Original samples Recognition Tar1 Tar2 Tar3 Tar4 Tar5 Tar1 48 1 1 0 0 Tar2 0 50 0 0 0 Tar3 0 1 48 0 1 Tar4 1 0 0 49 0 Tar5 0 0 1 0 49 Table 1. Recognition results of the proposed method on the original sample
-
噪声干扰是图像处理领域的一个常见问题,也是需要不断克服的难点问题。对于红外图像目标识别问题,当测试样本的信噪比远低于训练样本时,两者之中的目标特性将会存在较大的差异,导致训练得到的分类器性能下降。为在噪声干扰条件下进行文中方法的实验,首先按照参考文献[19]中的相关思路进行噪声测试集的构造。按照设定的信噪比进行噪声生成,并与原始红外图像进行混合处理,获得特定噪声水平的红外图像。在此条件下,文中对提出方法和4类对比方法进行分别测试,统计各类方法的平均识别率如图3所示。从总体趋势来看,噪声干扰对各类方法的性能均产生了较为显著的影响。对比信噪比为10 dB和−10 dB下的结果,都存在较大的差距。相比而言,文中方法的性能下降最为平缓,表明其受到噪声干扰的影响相对较小,显示其稳健性。文中通过SRC和CNN的有机融合,充分结合了两者的优势,对于噪声干扰的适应性得到了进一步的增强。
-
原始测试样本中,目标都是完整存在的,因此可以通过目前的全面特性进行训练和分类。然而,图2中的车辆目标均可能发生部分遮挡,导致获取图像中的目标是不完整和部分缺失的。该实验中,首先基于原始样本进行遮挡样本的模拟。具体地,以完整目标的区域为参照,采用背景像素对其局部区域进行填充处理。根据填充区域的比例定义不同的遮挡程度。在获得遮挡测试集的条件下,对各类方法进行测试,统计它们的平均识别率如表2所示。随着遮挡水平的不断提升,各类方法的性能下降十分明显。在各个遮挡水平下进行横向比较,可以看出文中方法均可以保持最高的平均识别率,表明其稳健性。与噪声干扰的情况相近,文中通过结合SRC和CNN的优势,可以进一步提升识别方法对于遮挡样本的适应性。
Method Occlusion level 5% 10% 15% 20% 25% Proposed 95.4 89.7 80.3 72.4 61.8 SRC 94.2 86.9 79.1 68.2 58.6 CNN 93.6 85.8 77.8 66.8 57.2 JSRDeep 94.4 87.9 79.2 70.3 60.2 Texture 93.5 86.4 78.1 67.2 58.1 Table 2. Average recognition rates of different methods on the occluded samples
Infrared image target recognition method based on decision fusion of classifiers
doi: 10.3788/IRLA20210725
- Received Date: 2021-11-05
- Rev Recd Date: 2022-02-25
- Available Online: 2022-08-31
- Publish Date: 2022-08-31
-
Key words:
- infrared image /
- target recognition /
- decision fusion
Abstract: The problem of infrared image target recognition based on classifier decision fusion was proposed. The sparse representation-based classification (SRC) and convolutional neural network (CNN) were used as the basic classifiers. For the test sample, it was first classified based on SRC, and the reliability of the decision was judged based on the output decision variables. When it was determined that the recognition result is reliable, the recognition process ended and the target category was output. On the contrary, some candidate categories with higher confidence were selected according to the results of SRC, and CNN was employed to confirm the classification result in the next stage. In addition, the CNN output result and SRC were subjected to linear weighted fusion processing, and the final target category decision was made according to the fusion result. The proposed method integrated the advantages of both SRC and CNN classifiers to comprehensively improve the performance of infrared target recognition. At the same time, this hierarchical decision fusion method avoided the two classification processes for all samples, and could ensure the overall efficiency of the recognition algorithm. The experiment was carried out using five types of infrared images of common vehicle targets in daily life, and the original sample conditions, noise sample conditions and occlusion sample conditions were set respectively. By comparing with some existing methods, the results reflect the effectiveness and reliability of the proposed method.