Obținerea hărților de adâncime din imagini defocalizate folosind rețele neurale profunde / Deep depth from defocus for near range and in-situ 3D exploration

Monocular depth estimation is a challenging task in computer vision that has wide area of applications ranging from robotics to self-driving cars. Deep learning has revolutionized depth estimation, but prior techniques often neglect defocus blur, an important cue for depth estimation. Existing solutions typically require multiple images or focal stacks, but rarely work with single defocused images. To fill in this gap, we propose novel architecture called 2HDED:NET, that addresses both depth estimation and image deblurring from a single defocused image. Due to the absence of datasets containing naturally defocused images and depth ground truth, networks like 2HDED:NET are typically trained on synthetic data, a fact that reduces their performances on real data. In order to train 2HDED:NET to its full potential, we proposed a new dataset called iDFD containing naturally defocused images, double annotated with the all-in-focus image, and the Time-of-Flight depth map. Finally, the thesis explores the promising field of self-supervised learning, by converting 2HDED:NET to defocus map estimation in the absence of ground truth depth for training. To accomplish this, 2HDED:NET is enhanced with a defocus simulation module that reconstructs the defocused image from the all-in-focus one and the estimated defocus map. The model’s proficiency in defocus map estimation is on par with that of state-of-the-art supervised models that use multiple images. Comprehensive experiments conducted on various real or synthetic datasets validated the efficacy of the proposed approaches for depth or defocus map estimation in various settings, encompassing indoor and outdoor environments.