TransFuse: A Transform Learning Based Multisensor Fusion Framework

This letter presents a new transform learning (TL) based multisensor fusion framework referred to as TransFuse. Unlike the standard representation learning based techniques, TransFuse learns individual transforms for each sensor and fuses them using a common transform representation within a joint optimization formulation. Considering regression as a use case, both the non-kernelized and kernelized versions are presented; the solution steps for learning the transforms, coefficients, and the regression weights are provided. The performance of the proposed TransFuse is evaluated using two real-life datasets and comparisons with the standard well-known TL and dictionary learning techniques for regression are presented. The results demonstrate the superior performance of TransFuse compared to its counterparts and also show the importance of multisensor fusion.