Abstract | ||
---|---|---|
The number of pictures taken by smartphones is growing exponentially. However, the smartphones’ limitations both in size and cost negatively impact on the quality of the implemented sensors. At the same time, their computing power has also been steadily improving, allowing the usage of more complex processing methods to enhance images. In prior works, deep neural networks trained with matched sensor outputs and DSLR images have shown to bring substantial improvements to the images, compared to classical and handcrafted methods. We propose a lightweight attention-based network (LAN) that employs a convolutional layer to learn the input mosaic and an unsupervised pre-training strategy. Our method is validated on standard benchmarks and shown to improve over the state-of-the-art in both perceptual and fidelity terms without hindering GPU inference time on smartphone devices. Our code is available at: github.com/draimundo/LAN |
Year | DOI | Venue |
---|---|---|
2022 | 10.1109/CVPRW56347.2022.00096 | 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) |
Keywords | DocType | Volume |
smartphone devices,GPU inference time,unsupervised pre-training strategy,lightweight attention-based network,classical handcrafted methods,DSLR images,matched sensor,deep neural networks,complex processing methods,computing power,implemented sensors,cost negatively impact,smartphones,RAW-to-RGB smartphone image processing | Conference | 2022 |
Issue | ISSN | ISBN |
1 | 2160-7508 | 978-1-6654-8740-5 |
Citations | PageRank | References |
0 | 0.34 | 10 |
Authors | ||
3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Daniel Wirzberger Raimundo | 1 | 0 | 0.34 |
Andrey Ignatov | 2 | 30 | 6.66 |
Radu Timofte | 3 | 1880 | 118.45 |