基于眼动实验的分类视觉注意显著性预测仿真

Classification Visual Attention Saliency Prediction Simulation Based on Eye Movement Experiment

为了准确地预测出图像中吸引视觉的关键部分,提出基于眼动实验的分类视觉注意显著性预测方法。通过MSRA10K图像库训练全卷积神经网络,得到图像的初步显著性区域特征。对其进行超像素优化,提取多尺度图像特征,对比局部融合颜色和全局颜色,形成低层特征显著图。提取各个图像块的主成分,计算主成分空间中图像块的局部以及全局可区分性,获取模式显著图。引用空间离散度度量分配相应的权重,将两者进行融合,准确预测分类视觉注意显著性区域。将所提方法与较为经典的两种方法进行实验对比,实验结果表明,所提方法能够更加准确预测出图像中的显著性区域。

In order to accurately predict the key parts of image attracting vision, this article puts forward a method to predict the classification visual attention saliency based on eye movement experiment. The whole convolution neural network was trained by MSRA10 K image database, so that the preliminary saliency region features of image could be obtained. Based on super-pixel optimization, multi-scale image features were extracted. After local fusion color and global color were compared, low-level feature saliency map was formed. Moreover, the principal components of each image block were extracted. Meanwhile, the local distinguish ability and global distinguish ability of image block in the principal component space was calculated to get the pattern saliency map. The spatial dispersion measure was used to allocate the corresponding weights. Based on the fusion result, the visual attention saliency regions were accurately predicted. Finally, the proposed method was compared with the classical methods. Simulation results show that the proposed method can predict the saliency region in image more accurately.