Single Image Super Resolution Based On Neural Networks For Text And Face Recognition

This thesis is focussed on super-resolution (SR) methods for improving automatic recognition system (Optical Character Recognition, face recognition) in realistic contexts. SR methods allow to generate high resolution images from low resolution ones. Unlike upsampling methods such as interpolation, they restore spatial high frequencies and compensate artefacts such as blur or jaggy edges. In particular, example-based approaches learn and model the relationship between low and high resolution spaces via pairs of low and high resolution images. Artificial Neural Networks are among the most efficient systems to address this problem. This work demonstrate the interest of SR methods based on neural networks for improved automatic recognition systems. By adapting the data, it is possible to train such Machine Learning algorithms to produce high-resolution images. Convolutional Neural Networks are especially efficient as they are trained to simultaneously extract relevant non-linear features while learning the mapping between low and high resolution spaces. On document text images, the proposed method improves OCR accuracy by +7.85 points compared with simple interpolation. The creation of an annotated image dataset and the organisation of an international competition (ICDAR2015) highlighted the interest and the relevance of such approaches. Moreover, if a priori knowledge is available, it can be used by a suitable network architecture. For facial images, face features are critical for automatic recognition. A two step method is proposed in which image resolution is first improved, followed by specialised models that focus on the essential features. An off-the-shelf face verification system has its performance improved from +6.91 up to +8.15 points. Finally, to address the variability of real-world low-resolution images, deep neural networks allow to absorb the diversity of the blurring kernels that characterise the low-resolution images. With a single model, high-resolution images are produced with natural image statistics, without any knowledge of the actual observation model of the low-resolution image.