为了提升对光学遥感图像中弱小运动目标的检测能力, 提出一种基于鹰眼视网膜视觉系统层次结构的运动检测方法。首先, 基于鹰眼视网膜的分层特性, 结合各层主体细胞的生理结构与功能, 构建各层相应的滤波器, 抑制背景微位移和杂散噪声; 然后, 在Reichardt运动检测模型的基础上增加时域高通滤波与ON-OFF双通道滤波来估计目标运动矢量, 这样不仅能克服传统Reichardt运动检测器对阶跃边界响应复杂, 而且能有效增强运动检测的敏感性; 最后利用高级视觉神经系统的分层特点, 以空域相似度大小为基准进行多尺度映射与运动矢量显著图融合, 构建多尺度处理精细检测运动特征。试验结果表明, 本文算法的平均信杂比改善为56.20 dB, 正确率为99.71%, 综合评价指标F1值为3.63e-02, 相较于传统Reichardt模型的F1值提升了27.82%。本文方法较传统运动检测算法不仅能提高复杂背景的干扰抑制性能, 而且能显著提升小目标小位移的检测能力。

To improve the detection ability of weak and small moving targets in optical remote sensing images, a motion detection method based on the hierarchical structure of the eagle-eye retinal vision system was proposed. Firstly, based on the stratification characteristics of the eagle-eye retina, combined with the physiological structure and function of the main cells of each layer, corresponding filters of each layer were constructed to suppress the background micro-displacement and spurious noise. Then, based on the Reichardt motion detection model, time domain high-pass filtering and ON-OFF dual-channel filtering were added to estimate the target motion vector, which overcomes the complex response of the traditional Reichardt motion detector to the step boundary and also effectively enhances the sensitivity of motion detection. Finally, using the hierarchical characteristics of the advanced visual nervous system, multi-scale mapping and the motion vector saliency map were combined based on the degree of spatial similarity, and multi-scale processing was used to detect the motion features. The experimental results show that the average signal-to-noise ratio of the proposed algorithm is improved to 56.20 dB, the correct rate is 99.71%, and the comprehensive evaluation index F1 is 3.63e-02, which is 27.82% higher than that of the traditional Reichardt model. Compared with the traditional motion detection algorithm, the proposed method can improve the interference suppression performance of complex background and also enhance the detection ability of a small target and small displacement significantly.