Spatially-adaptive filter units for compact and efficient deep neural networks
- Submitting institution
-
The University of Birmingham
- Unit of assessment
- 11 - Computer Science and Informatics
- Output identifier
- 91987496
- Type
- D - Journal article
- DOI
-
10.1007/s11263-019-01282-1
- Title of journal
- International Journal of Computer Vision
- Article number
- -
- First page
- 2049
- Volume
- 128
- Issue
- 8-9
- ISSN
- 0920-5691
- Open access status
- Compliant
- Month of publication
- January
- Year of publication
- 2020
- URL
-
-
- Supplementary information
-
-
- Request cross-referral to
- -
- Output has been delayed by COVID-19
- No
- COVID-19 affected output statement
- -
- Forensic science
- No
- Criminology
- No
- Interdisciplinary
- No
- Number of additional authors
-
2
- Research group(s)
-
-
- Citation count
- 1
- Proposed double-weighted
- No
- Reserve for an output with double weighting
- No
- Additional information
- The paper proposes a new convolutional filter composed of displaced aggregation units (DAU) as a novel elementary unit for deep neural networks. The main novelties comprise decoupling the parameter count from the receptive field size and automatically adjusting the filter’s spatial focus. This is significant as it substantially reduces the number of parameters with respect to conventional networks and allows automatic adaptation to specific computer tasks, as demonstrated on classification, semantic segmentation, and blind image deblurring. DAUs provide a seamless substitution of convolutional filters in existing state-of-the-art architectures, such as ResNet101, DeepLab (the code is publically available https://github.com/skokec/DAU-ConvNet).
- Author contribution statement
- -
- Non-English
- No
- English abstract
- -