用于显著性检测的除法归一化方法

A Division Normalization Method for Saliency Detection

针对现有的显著性检测方法输出分辨率低和缺乏生物合理性的问题,提出一种既有生物合理性,又能够获得全分辨率显著图的除法归一化方法.首先将L*a*b*颜色空间的输入图像分解成绿、红、蓝、黄和亮度5个特征通道,然后利用各通道的能量将每个通道进行归一化处理,再将5个归一化的通道在L*a*b*颜色空间中进行合成,最后利用欧几里得范数计算得到显著图,该过程模仿了初级视觉皮层中同类特征的相互抑制作用.实验结果表明,该方法计算简单高效,在心理物理学模板测试和人眼注视点预测上优于其他经典的显著性检测方法.

Since most existing approaches of saliency detection output low-resolution and biologically im-plausible saliency maps, in this paper, we propose a division normalization method to solve these problems. Our method is biologically motivated and is capable of producing a full resolution saliency map. Firstly, we decompose the input image fromL*a*b* color space into five feature channels: green, red, blue, yellow and luminance. Then, we normalize the channels by use of their respective energy. Next, we integrate five nor-malized channels in theL*a*b* color space. Finally, we employ the Euclidean norm to compute the saliency map. This procedure simulates the same feature suppression in primary visual cortex of humans. Experi-mental results show that the method is simple and computationally efficient, and outperforms other conven-tional approaches of saliency detection both on the psychophysical pattern test and on the eye fixations pre-diction tasks.