基于数据挖掘的立体视觉效果评价方法研究

Research on Stereo Visual Effect Evaluation Method Based on Data Mining

为了提升三维图像立体视觉效果评价质量,研究基于数据挖掘的立体视觉效果评价方法。利用二维Gabor滤波器提取图像的单目特征,即自然场统计特征;按照双目竞争特点融合左右图像,获取独眼图,提取双目特征,即方向梯度直方图特征;级联两个特征,获取左、右图像特征,归一化处理左、右图像特征的局部亮度系数,结合主观立体视觉效果评价分数,输入改进卷积神经网络,训练网络,获取客观立体视觉效果评价模型,在模型内输入测试图像,输出评价分数。实验证明,上述方法可有效评价三维图像的立体视觉效果;卷积核数量为50个,所提方法的评价质量最佳;在评价不同失真类型三维图像时的皮尔森线性相关系数、斯皮尔曼等级相关系数均较高,具备较优的三维图像立体视觉效果评价质量。

In order to improve the evaluation quality of stereoscopic visual effects of 3D images,a method for e⁃valuating stereoscopic visual effects based on data mining is studied.Use the two-dimensional Gabor filter to extract the monocular features of the image,that is,the statistical features of the natural field;According to the characteris⁃tics of binocular competition,the left and right images are merged,the cyclops are obtained,and the binocular fea⁃tures are extracted,that is,the directional gradient histogram feature;Concatenate two features to obtain left and right image features,normalize the local brightness coefficients of left and right image features,combine subjective stereo vision effect evaluation scores,input improved convolutional neural network,train the network,and obtain ob⁃jective stereo vision effect evaluation model,in which test images are input and evaluation scores are output.Experi⁃ments show that this method can effectively evaluate the stereoscopic effect of 3D image.The number of convolution kernel is 50,and the evaluation quality of this method is the best.Pearson's linear correlation coefficient and Spearman's rank correlation coefficient are both high when evaluating 3D images with different distortion types,and the quality of stereo visual effect evaluation of 3D images is better.