Title:
MULTI-BANDWIDTH SEPARATED FEATURE EXTRACTION CONVOLUTION LAYER FOR CONVOLUTIONAL NEURAL NETWORKS
Document Type and Number:
WIPO Patent Application WO/2022/073408
Kind Code:
A1
Abstract:
Methods, processing units and media for multi-bandwidth separated feature extraction convolution in a neural network are described. A convolution block splits input channels of an activation map into multiple branches, each branch undergoing convolution at a different bandwidth by using down-sampling of the inputs. The outputs are concatenated by up-sampling the outputs of the low-bandwidth branches using pixel shuffling. The concatenation operation may be a shuffled concatenation operation that preserves separated multi-bandwidth feature information for use by subsequent layers of the neural network. The methods also apply frequency-based and magnitude-based attention to the weights of the convolution kernels based on the frequency band locations of the weights.
Inventors:
QUADER NIAMUL (CA)
KHALIL MD IBRAHIM (CA)
LU JUWEI (CA)
DAI PENG (CA)
LI WEI (CA)
KHALIL MD IBRAHIM (CA)
LU JUWEI (CA)
DAI PENG (CA)
LI WEI (CA)
Application Number:
PCT/CN2021/117299
Publication Date:
April 14, 2022
Filing Date:
September 08, 2021
Export Citation:
Assignee:
HUAWEI TECH CO LTD (CN)
International Classes:
G06N3/08
Foreign References:
| US20200118002A1 | 2020-04-16 | |||
| US20200242734A1 | 2020-07-30 | |||
| US20200265106A1 | 2020-08-20 | |||
| CN109903301A | 2019-06-18 | |||
| US20200042871A1 | 2020-02-06 |
Download PDF: