Model-Based Image Signal Processors via Learnable Dictionaries

Digital cameras transform sensor RAW readings into RGB images by means of their Image Signal Processor (ISP). Computational photography tasks such as image denoising and colour constancy are commonly performed in the domain, part due to inherent hardware design, but also appealing simplicity noise statistics that result from direct readings. Despite this, availability is limited comparison with abundance diversity available data. Recent approaches have attempted bridge this gap estimating mapping: handcrafted model-based methods interpretable controllable usually require manual parameter fine-tuning, while end-to-end learnable neural networks large amounts training data, at times complex procedures, generally lack interpretability parametric control. Towards addressing these existing limitations, we present a novel hybrid data-driven ISP builds on canonical operations both interpretable. Our proposed invertible model, capable bidirectional mapping between domains, employs learning rich representations, i.e. dictionaries, free supervision additionally enable simple plausible data augmentation. We evidence value our generation process extensive experiments under reconstruction tasks, obtaining state-of-the-art performance both. Additionally, show can learn meaningful mappings few samples, models trained dictionary-based augmentation competitive despite having only or zero ground-truth labels.