基于中央凹图像显著性和扫视倾向的注视点转移预测模型

Scanpath Estimation Based on Foveal Image Saliency and Saccadic Bias

注视点转移预测是图像显著性建模中的一个重要研究领域.现有算法大多过于繁琐,并且局限于利用单张静态显著图来预测,很少能考虑到注视点转移是一个动态的过程.针对以上问题,我们提出一种具有生物依据的注视点转移的预测方法,利用了3大因素:视网膜中央凹在显著目标选取中的动态预测作用、注视点移动在距离和方向上的倾向以及短期记忆对注视点返回的抑制机制.通过三者得到的联合转移概率随机产生候选点,从而逐点生成扫视轨迹.相比于其他模型,所提议方法在客观的衡量标准下,在多个数据库中能更准确、更高效地预测扫视路径.

The estimation of gaze shifting has been an important research area in saliency modeling. Most of the existing methods tend to be complex in computation and are limited to estimating scanpaths within only one saliency map, while the gaze movement is a dynamic progress. To solve the above problems, a novel bio-inspired method for predicting eye movements is proposed. There are three principal factors： the effect of foveal images in finding salient regions dynamically, the saecadic bias in the distance and direction of gaze shifts, and the mechanism of IoR（Inhibition of Return） in short-term memory. Based on the probability map from the three factors, we can randomly generate candidates of the next fixation and get the final scanpath point by point. Compared to existing models, our method performs more accurately and efficiently in several datasets under the evaluation of objective comparison measures.