Nonlinear compressive imaging for machine learning
Published : 10 January 2019
In a context where the deployment of image sensors combined with computer vision tend to grow very quickly, the major challenges lie in information and signal processing. In the field of smart low-power sensors, the emerging breakthrough technology named Compressive Sensing is of major interest. In the case of embedded systems, autonomous decision-making becomes one of the core device feature while available resources (i.e., memory load, computing complexity and power consumption) remain highly limited. Indeed, the power consumption due to the sensor with dedicated signal processing is largely related to the overall data bandwidth and involved signal dimensionality. In particular, recent theoretical results demonstrate that standard Machine Learning approach can be advantageously applied in the compressed signal domain. However, those results are only restricted to the methods said as « linear », i.e. based on linear projections. The first objective of this PhD will thus be to properly identify theoretical limitations related to the combination of advanced Machine Learning with Compressive Sensing. It will aim at providing cutting-edge algorithm principles outperforming state-of-the-art tradeoffs between resources and inference accuracy. Thanks to a solid background in the laboratory on these fields of research, the goal of this thesis will be to evaluate the interest of introducing non-linearity during the acquisition process in order to improve the overall efficiency. This will help to define proper levers for smart sensor design enabling close-to-sensor context recognition (e.g., specific object detection with a highly limited hardware).