Multi-view Remote Sensing Tensor Canonical Correlation Analysis Networks for Scene Recognition

PROJECT TITLE :

Tensor Canonical Correlation Analysis Networks for Multi-view Remote Sensing Scene Recognition

ABSTRACT:

It has been demonstrated that using a convolutional neural network, also known as CNN, is an efficient method for automatically extracting high-level features from remote sensing (RS) images. Many different iterations of the CNN model, such as the principal component analysis network (PCANet), the canonical correlation analysis network (CCANet), the multiple scale CCANet (MS-CCANet), and the multiview CCANet, have been proposed (MCCANet). The PCANet is optimized for the abstraction of features from a single view, whereas in many real-world practices, the RS data are frequently observed from a large number of different perspectives. Although CCANet, MS-CCANet, and MCCANet can be used on data from two or more views, they only take into account the pair-wise correlation by calculating a series of two-order covariance matrices. This is the case even though these methods are applicable to more than two views. However, the high-order consistence has not yet been uncovered. This is because it is impossible to investigate it unless all perspectives are considered jointly and simultaneously. In this paper, we address this issue by putting forward a solution that we refer to as the tensor canonical correlation analysis network, or TCCANet. Specifically, TCCANet learns filter banks by simultaneously maximizing an arbitrary number of views with high-order correlation, and it solves the optimization problem by decomposing a covariance tensor. This allows it to learn filter banks more quickly. Following the convolutional stage, we move on to the binarization stage, where we employ block-wise histogram and binarization strategies to generate the final feature. In addition, we develop a Multiple Scale version of TCCANet, which we refer to as MS-TCCANet, in order to extract an enriched representation of the RS data by incorporating all of the preceding convolutional layers. This was done so that the model could better understand the data. The benefits of TCCANet and MS-TCCANet for RS scene recognition are illustrated by the results of numerical experiments performed on the RSSCN7 and SAT-6 datasets.

Did you like this research project?

To get this research project Guidelines, Training and Code... Click Here