基于极端学习机背景预测的红外小目标检测算法

Infrared small-target detection algorithm based on background prediction by extreme learning machine

为自适应检测复杂环境中的红外小目标,提出了基于极端学习机背景预测的红外小目标检测算法。首先,依据灰度值分布设计局部边缘敏感平滑滤波器,在相近的灰度范围内,使中心像素的灰度值等于邻域内多数灰度值的融合,对红外图像进行滤波,能够去除大量噪声并突出图像主要结构;其次,利用极端学习机对滤波后的图像建立回归模型,以邻域像素值为输入,以中心像素值为输出训练模型,并对背景进行预测,得到的图像与滤波后的图像做差,得到小目标显著图;最后,利用图像块对比特性对显著区域处理,使小目标区域均匀突出,抑制背景区域,并经过简单阈值操作,实现对红外小目标的检测。实验结果表明:与其他检测算法相比,在复杂背景下,本文算法检测结果的局部信噪比增益最高,单帧检测时间为0.18 s。本文算法对背景进行学习,发掘背景与目标的差异,提高了算法的适应能力,并且能够有效检测小目标。

In order to adaptively detect infrared small targets under complex background, an infrared smalltarget detection algorithm is proposed based on background prediction by extreme learning machine. Firstly, a local edge-sensitive smoothing filter is designed based on grayscale distribution, in which center pixel grayscale is defined by mixing together more pixels in neighborhood with similar grayscale. It can remove a large amount of noises and highlight main structure of infrared image after filtering the image by local edge-sensitive smoothing filter. Secondly, it establish a regression model for filtered image by using extreme learning machine, in which training is carried out by using the grayscale of pixels in neighborhood as input and the grayscale of center pixel as output. After training, the background is predicted by this model and subtracted from the filtered image to form a salient map for small targets. Finally, in order to unanimously highlight the region of small targets and restrain the background, the salient region is processed by using the contrast characteristics of image patches, and then small targets＇ detection can be realized by simple threshold operation. Experiment results show that, compared with other detection algorithms, the local signal-to-noise rate gain is the highest by the proposed algorithm under complex background, and the single-frame detection time is 0.18 s. The proposed algorithm can study the background and discover the difference between the background and the target, which improve the adaptability of the algorithm and can effectively detect small targets.