基于L_(1−2)时空域总变分正则项的红外弱小目标检测算法

Infrared small target detection via L_(1−2) spatial-temporal total variation regularization

针对红外图像序列中复杂背景干扰下容易出现的高虚警问题,提出一种基于L_(1-2)时空域总变分正则项的红外弱小目标检测算法。首先,将红外图像序列转化为时空域红外张量块,该步骤可利用张量的高维数据结构优势关联图像序列中的时空域信息。然后,利用加权Schattenp范数和L_(1-2)时空域总变分正则项对低秩背景成分进行重构,以保留背景中起伏剧烈的边缘和角点,提高稀疏目标的重构精度。最后,将目标张量恢复为图像序列,利用自适应阈值分割方法得到最终的目标图像。与另外5种检测算法进行对比实验,结果显示,该方法的虚警率较Maxemeidan算法、Tophat算法、LIRDNet算法、DNANet算法以及WSNMSTIPT算法平均分别下降了71.4%、71.1%、68.5%、74.3%和20.47%;而在检测实时性方面,该算法耗时为Maxemeidan算法、DNANet算法以及WSNMSTIPT算法的42.4%、82.9%和28.7%。实验结果验证了该方法在检测性能上的优越性,表明该算法能够显著提高复杂背景干扰下的目标检测精度和效率。

To solve the high false alarms caused by complex background clutters in infrared small-target detection,a novel detection method based on L_(1-2)spatial-temporal total variation regularization is proposed.First,the input infrared image sequence is transformed into a Spatial-Temporal Infrared Patch-Tensor(STIPT)structure.This step can associate the spatial and temporal information by using the high dimensional data structures in the tensor domain.Then,weighted Schatten p-norm and L_(1-2)spatial-temporal total variation regularization are incorporated to recover the low-rank background component to preserve the strong edges and corners,which can improve the accuracy of sparse target component recovery.Finally,the STIPT structure can be transformed into an infrared image sequence by the inverse operator,and an adaptive threshold segmentation is used to obtain the real target.The method is verified using a contrast test with other five methods,and the experimental results show that the false alarm rate by this method decreases to71.4%,71.7%,68.5%,74.3%and 20.47%compared with the Maxemeidan,Tophat,LIRDNet,DNANet and WSNMSTIPT algorithms.The time cost also decreased to 42.4%,82.9%and 28.7%of that of the Maxemeidan,DNANet and WSNMSTIPT.The extensive experimental results demonstrate the superiority of this method in detection performance,which can greatly improve the accuracy and efficiency of target detection with complex background clutters.