Deep Transform Learning for Multi-Sensor Fusion

This paper presents a Deep Transform Learning based framework for multi-sensor fusion. Deep representations are learnt for each of the sensors by stacking one transform after another. Subsequently, a common transform is utilized to fuse the deep representations of all sensors to estimate the output. Restricting to a regression use case, a joint optimization formulation is presented for learning the sensor-specific deep transforms, their coefficients, the common transform, its coefficient and the regression weights together. The requisite solution steps and the derivation of closed form updates for the transforms and associated coefficients are given. The performance of the proposed method is evaluated using two real-life datasets and comparisons with the state-of-the-art dictionary and transform learning techniques for regression are presented. Results show that the deep network has superior performance compared to other methods as it is able to learn the data representation more effectively than the other shallow variants. In addition to the multi-sensor case, estimation results with single sensors alone are also provided to demonstrate the importance of multi-sensor fusion.