多视角红外图像目标识别方法

随着红外传感器的性能提升和应用普及，获取同一场景下同一目标的多视角图像成为可能。为此，提出联合多视角红外图像的目标识别方法。首先对多视角红外图像进行聚类分析，获取多个视角子集。在每个视角子集中，红外图像具有较强的相关性。对于不同的视角子集，它们相对独立。为充分利用这种相关性和独立性，采用联合稀疏表示（JSR）对单个视角子集进行决策。特别地，对于只包含一个视角的子集，则直接采用经典的稀疏表示分类（SRC）进行处理。对于不同视角子集获取的决策结果，基于线性加权的思想进行融合处理，并根据融合后的决策变量判决多视角红外图像所属的目标类别。因此，所提方法在分析多视角红外图像内在关联性的基础上，分别对局部相关性和整体的独立性进行考察，并通过决策层的融合将两者融为一体，提高了最终决策的可靠性。实验中，在采集的多类交通车辆红外图像上进行识别，分别在原始图像、加噪声图像以及部分遮挡图像上对方法进行测试和验证，经过对比分析验证了提出方法的有效性。

Target recognition based on multi-view infrared images

With the improvement of the performance of infrared sensors and the popularization of applications, it becomes possible to obtain multi-view images of the same target in the same scene. Therefore, a target recognition method combining multi-view infrared images was proposed. First, the clustering analysis on multi-view infrared images was performed to obtain multiple view-view subsets. In each view subset, the infrared images shared high correlations. For different view subsets, they were relatively independent. In order to make full use of the correlation and independence, the joint sparse representation (JSR) was used to make decisions on single view subsets. In particular, for the subset with only one view, the classical sparse representation-based classification (SRC) was directly used for decision. For the decision results obtained by different view subsets, the fusion processing was carried out based on the idea of linear weighting. And the target category was determined according to the fused results. Therefore, on the basis of analyzing the inner correlation of the multi-view infrared images, the proposed method separately examined the local correlations and overall independence, and integrated them through the fusion on the decision-making layer, which improved the reliability of the final decision. Experiments were performed on the collected infrared images of multiple types of traffic vehicles. The proposed method was tested and verified on the original, noisy, and occluded samples. The effectiveness of the proposed method is verified by comparison with other methods.