基于纹理图像分析的生物视觉模型不变性评价

Invariance Evaluation for Biological Visual Model Based on Analysis of Textured Images

基于纹理图像,从计算机视觉角度对生物视觉模型——视皮层目标识别的标准模型进行定量分析与评价。对原始图像分别进行尺度、旋转及仿射等变化,利用标准模型提取变化后图像的生物视觉特征,再根据提取的生物视觉特征对纹理图像进行分类,采用图像分类结果的曲线下面积来定量分析和评价生物视觉模型是否具有不变性。大量与局部二元模式特征的对比实验表明,该模型提取的生物视觉特征对于纹理图像具备优良的尺度、旋转与仿射不变性。

From the view of computer vision, the particular quantitative analyses and evaluations about rotational, scale and affine invariance of standard model of object recognition in cortex are made based on textured images. The original textured images are scaled, rotated and affinely transformed respectively. Biological visual features of these transformed images, also called Standard Model Features（SMFs）, are extracted by the standard model. The SMFs are used to classify the images. After classification, the area under curve is utilized to quantitatively analyze and evaluate whether the standard model has invariance or not. Compared with the Local Binary Pattem（LBP） feature, a great deal of experiments show that the biological visual features extracted by the standard model have sunerior scale, rotational and affine invariance to textured images.