Prédiction de performance d'algorithmes de traitement d'images sur différentes architectures hardwares. (Image processing algorithm performance prediction on different hardware architectures)

In computer vision, the choice of a computing architecture is becoming more difficult for image processing experts. Indeed, the number of architectures allowing the computation of image processing algorithms is increasing. Moreover, the number of computer vision applications constrained by computing capacity, power consumption and size is increasing. Furthermore, selecting an hardware architecture, as CPU, GPU or FPGA is also an important issue when considering computer vision applications. The main goal of this study is to predict the system performance in the beginning of a computer vision project. Indeed, for a manufacturer or even a researcher, selecting the computing architecture should be done as soon as possible to minimize the impact on development. A large variety of methods and tools has been developed to predict the performance of computing systems. However, they do not cover a specific area and they cannot predict the performance without analyzing the code or making some benchmarks on architectures. In this works, we specially focus on the prediction of the performance of computer vision algorithms without the need for benchmarking. This allows splitting the image processing algorithms in primitive blocks. In this context, a new paradigm based on splitting every image processing algorithms in primitive blocks has been developed. Furthermore, we propose a method to model the primitive blocks according to the software and hardware parameters. The decomposition in primitive blocks and their modeling was demonstrated to be possible. Herein, the performed experiences, on different architectures, with real data, using algorithms as convolution and wavelets validated the proposed paradigm. This approach is a first step towards the development of a tool allowing to help choosing hardware architecture and optimizing image processing algorithms.