-
采用中波红外目标数据集(MWIR)开展实验,覆盖波长范围3~5 μm,其中包含的10类目标类别及图像示意图如图3所示。该数据集采用目标圆周运动方式进行全方位采集,从而有效覆盖不同方位条件下的红外图像数据。实验中,对于任一类目标,通过图像平移、裁剪(包含目标区域)的方式获得120幅图像用作训练,100幅图像用于测试(测试样本与训练样本不同)。
从公开文献中的红外目标识别算法中选取4类进行对比实验,包含采用SVM、SRC、JSR和CNN的方法。其中,SVM[10]和SRC[11]方法对红外目标图像提取的PCA特征矢量进行分类;JSR方法采用与文中一致的联合表征模型[16],但其特征均为投影变换类;CNN方法采用与文中相同的网络结构,但直接以原始图像为输入进行训练和决策,不涉及深度特征的进一步处理。
后续实验首先基于原始MWIR图像样本开展,然后对测试样本进行不同程度的噪声添加,检验所提方法的噪声稳健性,最后,对训练集进行不同程度的削减,考核方法在少量训练样本条件下的识别能力。
-
首先对3.1节设置的原始测试样本进行测试。此时,测试样本和训练集采集自相近的场景,相似度较高,目标识别难度相对较低。图4显示了所提方法对图3所示10类目标原始测试样本的混淆矩阵,对角线对应不同目标的正确识别率。定义平均识别率为
${P_{{\rm{av}}}}{\rm{ = }}{N_{\rm{c}}}/{N_{\rm{T}}}$ ,式中${N_{\rm{c}}}$ 为所有类别正确识别的测试样本总数;${N_{\rm{T}}}$ 为所有测试样本的总数,据此计算得到所提方法的平均识别率为97.7%。表1对比了所提方法与4类对比方法的识别结果。对比可见,所提方法取得了最高的平均识别率,表明其对于MWIR数据集中原始测试样本的识别有效性。与JSR方法相比,文中通过运用鉴别力更强的深度特征有效提升了最终性能。与CNN方法相比,文中通过对其学习的多层次深度特征的联合运用,进一步提升了识别性能。上述结果验证了所提方法的有效性及其优势性能。Method Average recognition rate Proposed 97.7% SVM 94.5% SRC 95.1% JSR 96.6% CNN 97.2% Table 1. Comparison of recognition results on the original test samples with different methods
-
在原始测试样本的条件下,测试样本与训练集获取条件相近,具有相当的信噪比(SNR)。实际过程中,测试样本可能采集自不同的场景,受到不同程度的噪声干扰。为此,文中基于原始测试样本构造不同信噪比下的噪声样本。具体地,根据测试样本的能量,按照预设的信噪比要求构造加性高斯白噪声,并将其加入当前测试样本。通过调整信噪比要求,即可获得不同噪声水平(含信噪比10 dB、5 dB、0 dB、−5 dB和−10 dB)下的测试集。采用所提方法和对比方法对噪声样本进行测试,获得如图5所示的统计结果。可见,随着测试样本信噪比的降低,各类方法均出现性能下降。对比来看,所提方法在各个信噪比均取得了最高的平均识别率。特别是,CNN方法在低噪声水平下性能下降十分剧烈,文中方法引入JSR作为末端分类器提升了整体噪声稳健性。上述结果验证了所提方法对于噪声干扰的有效性。
-
训练样本的规模和数量是分类器识别性能的重要影响因素。实际过程中,考虑到非合作目标红外图像获取的难度,训练样本往往十分稀缺。为此,识别算法在小规模训练集条件下的适应性十分关键。为此,文中对原始训练集进行随机削减,获得80%、60%、40%和20%比例下的训练集。图6显示了不同方法在各个削减训练集下的平均识别率。可以看出,所提方法在各个比例下均可确定最高性能,显示其在少量训练样本条件下的有效性。CNN方法识别性能与训练样本规模紧密相关,这也导致其在低比例训练集条件下识别率下降十分显著。文中通过结合多层次深度特征和联合稀疏表示提高了对于少量训练样本的适应性。
Target recognition method of infrared imagery via joint representation of deep features
doi: 10.3788/IRLA20200399
- Received Date: 2020-10-19
- Rev Recd Date: 2021-01-20
- Available Online: 2021-05-12
- Publish Date: 2021-03-15
-
Key words:
- infrared imagery /
- target recognition /
- deep features /
- joint sparse representation
Abstract: For the target recognition of infrared imagery, a method was proposed via the combination of convolutional neural network (CNN) and joint sparse representation (JSR). CNN learned the deep features of the infrared target imagery, which described the multi-layer properties of the target. Different layers of deep features described the target charateristics from differnt aspects, so they can well complement each other. The joint use of multi-layer deep features could provide more valid information for target recognition. During the classification, JSR was employed to represent the multi-level deep feature vectors and the inner correlations among different features was used to improve the overall representation precision. Therefore, JSR not only made use of individual deep features but also considered their inner correlations. According to the outs from JSR, the target label of the input sample was determined based on the minimum error. The experiments were conducted based on mid-wave infrared (MWIR) dataset under the conditions of original test samples, noise test samples, and small training set. Simultaneously, the proposed method was compared with four previous methods. According to the experimental results, the proposed method achieves better performance under the three conditions, validating its potential in infrared imagery target recognition.