講演抄録/キーワード |
講演名 |
2019-10-18 10:45
[ショートペーパー]An attribution-based pruning method for single object detection network ○Rui Shi・Tianxing Li・Yasushi Yamaguchi(UTokyo) PRMU2019-34 |
抄録 |
(和) |
(まだ登録されていません) |
(英) |
Deep neural networks (DNNs) have achieved advanced results on different vision tasks. However, the cost of high computational complexity is not practical for real-time inferences running on mobile devices. In this paper, we propose an attribution-based pruning method for object detection network that reduce the computation while ensuring accuracy. First, Channel mask and spatial mask are designed to generalize attribution methods to detection networks for detecting convolution kernels that are firmly correlated with target output. Then, YOLOv3-tiny network is pruned using attribution maps and finetuned on an open-sourced mango dataset for evaluation. Compared to a state-of-the-art network trained with the same mango dataset, the experiment shows that our network achieves 83.4% computation reduction with only about 2.4% loss in accuracy. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Deep learning / Mango detection / Network pruning / Attribution methods / / / / |
文献情報 |
信学技報, vol. 119, no. 235, PRMU2019-34, pp. 17-20, 2019年10月. |
資料番号 |
PRMU2019-34 |
発行日 |
2019-10-11 (PRMU) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
PRMU2019-34 |