Thesis of Clément Peyrard

Super-resolution methods for objects classification and recognition in videos


The main challenge of this PhD is to explore classical and new methods for image and video super-resolution, in order to improve image-based technologies performances at Orange Labs. In Rennes (France), the MAS (Multimedia content Analysis and technologieS) team experiments audio and video-based techniques to extract relevant informations from various sources (TV, radio, internet broadcast) for indexation purposes. As some of these techniques already reach state of the art performances on standard sources, the idea of this work is to deal with the upper level of the processes: image quality and resolution. Super-resolution provides one with the ability to upscale (or “zoom”, “magnificate”) an image while reducing undesired artefacts such as blur, ringing or other noisy results that occur when using standard interpolation techniques. This subject has been and is still of great interest, as the call for high-definition images with the latest screen technologies or with the multiplication of camera-enabled mobile devices keeps on increasing. It gathers researchers from various imagery domains such as military, surveillance, medicine, astronomy or satellite. While “traditional” super-resolution is able to recover a high resolution image from multiple low resolution images by different approaches, or to smartly interpolate images content taking care of the high frequency content, recent methods based on learning frameworks have demonstrated their ability to outperform standard techniques. Therefore, one aim of this study is to explore the statistical and learning methods from object detection, recognition and classification such as neural networks, sparse representations, and evaluate the possible transposition that might be performed to super-resolution.

Advisor: Christophe Garcia