6533b85afe1ef96bd12b9ca0

RESEARCH PRODUCT

Development of new techniques for super-resolution video sequences : Towards a real-time implementation on Smart Camera

subject

description

These thesis works are part of an european project aiming to design a very hight resolution (8k) video camera. Within this project our team had the task of working on two technological aspects: (1) the design of a demonstrator carrying out a realtime deconvolution of a video stream coming from a very high resolution camera created by the consortium , (2) the design of a prototype allowing to increase the resolution and the level of detail of video streams from an input resolution of 4k to 8k using Super Resolution (SR) methods. This manuscript mainly presents the work related to the creation of the prototype realizing a Super Resolution method. In order to be able to assess the qualitative contribution of SR, a study on the relevance of the metrics is carried out in order to select the most suitable metrics for evaluating image quality improvement processes. Then, a new method of Spatial SR named LASSR for Local Adaptive Spatial Super Resolution was implemented to correct certain artefacts observed with a State of the Art SR method. Our method allows to locally adapt the processing according the texture, this adaptation is carried out automatically using a decision model generated during supervised learning. Following the validation of our method by a panel of experts during a psychovisual assesment, the IP module was developed to operate with very large volumes of data on FPGA targets and to operate in real time. A toolbox of components has been created in order to describe the module in a generic way. This allows the LASSR module to be adapted to different configurations (image size, pixel dynamics . . .). The major element of this toolbox is the generic 2-D convolution filter : its size and its coecients can be tuned (the FPGA chip limits this filter). Encouraging results have been obtained regarding the increase in resolution for a 4k grayscale output image, it is possible for a single LASSR module to achieve an output frequency of 17.03 fps. On the FPGA of the Xilinx VC707 development board, it is theoretically possible to reach a frequency of 136.2 fps by parallelizing 8 LASSR modules.