Learning object recognition based on compressive sampling

Abstract

In this paper, we propose an object recognition algorithm allowing for learning in the compressed space and reconstructing to the spatial space when images need to be processed in their spatial forms. Instead of using the traditional cameras, a novel compressive sampling camera is simulated to directly capture the natural scene to compressed images based on compressive sampling theory. We evaluate the recognition performance and reconstruction quality on a traffic database providing a solution to the reliable situation awareness problem for the self-driving cars. We also exploit the effectiveness of our method on a publicly available face database. It is experimentally observed that the proposed approach can obtain a high recognition rate and achieve the image reconstruction.