基于加权相似性的MCCNN训练集选择方法

MCCNN training set selection method based on weighted similarity

为解决MCCNN网络立体匹配的训练数据集选择问题,研究一种基于相关性比较、余弦相似性和结构相似性的加权度量选择方法,通过实验确定三者的加权系数,使用三者的加权值衡量训练集与待匹配图像数据分布的互相似性、训练集本身的自相似性,以互相似性和自相似性加和值最高的对应数据集作为选择的训练集。通过InStereo2k图像和实拍图像实验,验证了该方法的有效性。

To solve the problem of selecting the training data set for the stereo matching of the MCCNN network,a weighted metric selection method based on correlation comparison,cosine similarity and structural similarity was proposed.The weighting coefficients of the three were determined through experiments.The weighted values of the three were used to measure the mu-tual similarity between the training set and the data distribution of the image to be matched,as well as the self-similarity of the training set itself.The corresponding data set with the highest sum of mutual similarity and self-similarity was selected as the training set.The effectiveness of the method was verified through experiments with InStereo2k data sets and real pictures.