The study introduces a deep learning method for focus stacking – a technique used to create all-in-focus images from multiple frames with different focus planes. Traditional deep learning approaches have limitations with real-world imagery, often being trained on small, low-res data. This study introduces a new dataset of 94 high-res bursts of raw images with focus bracketing. The new deep learning algorithm trained on this data showed comparable results with existing commercial solutions while being significantly more tolerant to noise.

 

Publication date: 29 Nov 2023
Project Page: https://github.com/araujoalexandre/FocusStackingDataset
Paper: https://arxiv.org/pdf/2311.17846