【摘要】 根据人类视觉系统的特点,提出一种自适应归一化图像分辨率的预处理方法。该方法通过计算图像灰度分布情况,根据图像所包含显著区域大小的不同,自适应的确定图像分辨率,打破了谱残差算法无论什么图像均将其归一化为固定分辨率大小(64×64)的局限;同时该算法还能够根据图像中所包含目标的不同大小,完整地提取出所有显著性目标,有效克服了谱残差算法倾向于小目标检测的缺点。实验结果表明:相对于谱残差算法,新算法简单,耗时少于0.1 s,检测准确率提高约15%,可以得到显著区域更为全面的检测效果。更多还原

【Abstract】 According to the characteristics of human visual system,an adaptive normalizing method for preprocessing the image resolution is proposed. We analyze the size of the saliency region by calculating the gray distribution of the input image; then we define the image resolution adaptively. This method can help to do away with the limitation of normalizing all images into a fixed resolution( 64×64) and to avoid the usually committed error of underestimating the total size of all effectively saliency regions when using the spectral residual method. The experimental results on natural images and their analysis show preliminarily that this approach can:( 1) improve the detection accuracy by consuming the time of less than 0. 1 second,thus improving hit rate by 15%;( 2) estimate correctly the total size of all saliency regions.更多还原