基于偏振权重局部对比度的目标检测

Object detection based on polarization-weighted local contrast method

分焦平面红外偏振探测器输出的是红外偏振马赛克图像,传统处理流程需要进行去马赛克恢复出四个偏振通道的完整图像,然后再实现后续的任务。然而,去马赛克过程会引入误差,而且计算复杂度高、耗时长。针对如何直接在红外偏振马赛克图像上进行目标检测的问题,本文提出了一种偏振权重局部对比度的目标检测方法。首先分析了目标与背景的偏振特性差异;然后设计了红外偏振马赛克图像的斯托克斯矢量计算卷积核;在此基础上提出了基于偏振权重的偏振度显著图,在偏振度显著图上利用自适应阈值操作实现目标检测。此外,利用边缘检测方法进一步优化目标检测结果,得到更加完整的检测结果。最后,使用采集的红外偏振马赛克数据集验证了所提出的目标检测算法在复杂背景以及恶劣天气影响下的鲁棒性。

The images captured by the Division-of-Focal-Plane(DoFP)infrared polarimeters have checkboard effect.Thus,a polarization demosaicking processing of DoFP images is demanded to recover the full resolution polarization images,based on which,the subsequent tasks are then performed.However,the demosaicking processing is usually timeconsuming and may introduce demosaicking errors.To achieve object detection by directly using infrared polarization DoFP image,a polarization-weighted local contrast object detection method is proposed.The difference of polarization characteristics between the object and background is first analyzed.Then,a convolution kernel is designed to calculate the Stokes vector directly from original infrared polarization DoFP images.A polarization-weighted saliency map of the degree of polarization image is also proposed,which is used for object detection with the adaptive thresholding.In addition,an edge detection method is used to refine the target detection results and obtain more complete detection results.The experiment results on the infrared polarization DoFP images dataset demonstrate that the proposed object detection algorithm is robust to the conditions of complex background and bad weather.