Feature Correlation Analysis of Two-Branch Convolutional Networks for Multi-Source Image Classification

With the development of multi-sensor imaging technology, constructing a multi-branch network model has become an important requirement for fusion decision. In the literature, many two-branch networks are proposed to interpret multisensor data and also get the satisfactory results. In this paper, we divide these models into two types and study them by mining changes in data and features. The main method used is analysis the correlation between the features of the same layer from the first branch and the second branch. The task of dual-source image classification serves as a means of experimentation. When the classification network is trained, the features of each layer are extracted for experimental analysis. Extensive experiments and analysis on the dataset IEEE_grss_dfc_2017 show that the analysis is meaningful. It is found that the correlation of the features from the low level to the high level is more and more consistent, a quantitative analysis is given in this paper.