Single-image blind motion deblurring : Bridging blur formation models with data-driven learning.

This thesis addresses the challenging, ill-posed problem of motion deblurring, a prevalent source of image degradation affecting numerous applications, such as photography, medical imaging, and robotics. Current state-of-the-art deep learning-based deblurring networks, while demonstrating impressive performance on specific datasets, often struggle with generalization to real-world scenarios due to their reliance on learning directly from blurry/sharp image pairs. Conversely, while often generalizing better under their assumed conditions, classical modelbased approaches frequently utilize overly simplified blur models (e.g. uniform blur), limiting their accuracy. This thesis proposes novel hybrid approaches that synergistically combine the strengths of model-based and data-driven methods. Our first contribution is a motion blur Kernel Prediction Network (KPN), which characterizes non-uniform motion blur using a set of image-adaptive basis kernels and their mixing coefficients.This efficient representation overcomes the limitations of simplistic blur models, enabling accurate modeling of complex, real-world blur patterns. The effectiveness of KPN is demonstrated in diverse datasets and settings. Specifically, the KPN is integrated into two distinct blind deblurring algorithms: one employing an adaptation of the Richardson-Lucy deconvolution algorithm and another leveraging a recent non-blind restoration network. The latter leads to the Joint Motion Kernel Prediction and Deblurring (J-MKPD) blind restoration method, a joint training approach that combines the motion blur KPN with an unrolled plug-and-play restoration network. Motivated by the high-quality results obtained with J-MKPD, this thesis further explores alternative degradation models that are better suited for specific scenarios. A camera trajectory prediction network is developed to estimate the motion blur kernel field resulting from camera shake. This network, when jointly trained with the same unrolled plug-and-play restoration network as J-MKPD, forms the Joint Motion Trajectory Prediction and Deblurring (J-MTPD) method, specifically designed for camera shake blur. A third method, Joint Motion Offsets Prediction and Deblurring (J-MOPD), utilizes a more expressive offset-based representation of the motion blur kernel field. While more challenging to train, this representation offers increased flexibility. A novel training strategy is introduced to constrain the solution space, enabling efficient integration with the unrolled plug-and-play network and resulting in a blind deblurring method assuming locally uniform blur. Recognizing the crucial role of training data in the generalization capabilities of end-to-end deblurring networks, this thesis comprehensively analyzes existing datasets, identifying their limitations and the underlying causes of poor generalization. Based on this analysis, a novel procedure is proposed to generate synthetic training data. This methodology focuses on simulating how blur manifests in realworld scenarios, allowing the generation of a virtually unlimited supply of diverse, high-quality training pairs. The resulting J-MKPD, J-MTPD, and J-MOPD methods constitute a suite of publicly available deblurring algorithms demonstrating excellent cross-dataset performance, especially on real-world image datasets. J-MKPD proves effective for motion blur up to 33 pixels, while J-MTPD excels in handling camera shake blur, and J-MOPD is well suited for locally uniform blur. Furthermore, the thesis explores the application of J-MKPD for super-resolution and demonstrates the video generation capabilities of J-MTPD and J-MOPD. This work bridges the gap between model-based and data-driven approaches, providing robust solutions while offering valuable insights into the limitations of existing techniques and the challenges in creating truly generalizable deblurring methods.