The high resolution of modern cameras puts significant performance pressure on image processing pipelines. Tuning their many parameters for speed is subject to stringent image quality constraints and requires significant efforts from skilled programmers. Because quality is driven by perceptual factors, with which most quantitative image metrics correlate poorly, developing new pipelines involves long iteration cycles, alternating between software implementation and visual evaluation by human experts. These concerns are compounded on modern computing platforms, which are increasingly mobile and heterogeneous.
This dissertation demonstrates that differentiable algorithms, together with carefully curated training datasets and improved image quality metrics, promise renewed and scalable progress for image processing. We employ supervised learning towards the design of high-performance, high fidelity algorithms whose parameters can be optimized automatically on large-scale datasets via gradient-based methods. We present applications to low-level image restoration and real-time image enhancement on mobile devices.
Thesis Supervisor: Prof. Frédo Durand