2020
DOI: 10.1109/tpami.2019.2907634
|View full text |Cite
|
Sign up to set email alerts
|

Towards Efficient U-Nets: A Coupled and Quantized Approach

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
1
1
1

Citation Types

1
49
0

Year Published

2020
2020
2022
2022

Publication Types

Select...
4
1
1

Relationship

0
6

Authors

Journals

citations
Cited by 51 publications
(50 citation statements)
references
References 38 publications
1
49
0
Order By: Relevance
“…We therefore thought to optimize the U-Net itself, with features known from different previous works, and to systematically investigate each part of the network. Improving U-Nets in terms of efficiency is rarely performed, compared to accuracy, but has been proposed recently by coupling stacked U-Nets for landmark detection together with evaluating several quantization approaches [20] or to use a very reduced U-Net for low power satellite segmentation [21].…”
Section: Related Workmentioning
confidence: 99%
“…We therefore thought to optimize the U-Net itself, with features known from different previous works, and to systematically investigate each part of the network. Improving U-Nets in terms of efficiency is rarely performed, compared to accuracy, but has been proposed recently by coupling stacked U-Nets for landmark detection together with evaluating several quantization approaches [20] or to use a very reduced U-Net for low power satellite segmentation [21].…”
Section: Related Workmentioning
confidence: 99%
“…1. The modified U-Net is formed by a so-called micro-block at each resolution level (instead of a single convolutional layer as in a conventional U-Net) 25,26 ; the output of each micro-block is further processed by a convolutional layer. As shown in Fig.…”
Section: B1 3d Modified U-net Modelmentioning
confidence: 99%
“…The purpose of these coupling connections is to promote reuse of the feature maps among the individual U-Nets, thereby reducing redundancy in learning among them. 25 Specifically, consider the n-th U-Net in the CU-Net model, n ¼ 1, ..., N, where N is the total number of U-Nets in the network. The input feature maps at its k-th micro-block, denoted by I n k , consist of both the feature maps generated within this U-Net itself and that generated from all its proceeding U-Nets.…”
Section: B2 3d Coupled U-net (Cu-net) Modelmentioning
confidence: 99%
See 2 more Smart Citations