TY - GEN
T1 - Manipulating Neural Network Block for Robust Image Segmentation
AU - Kim, Hyungjoon
AU - Kim, Hyeonwoo
AU - Cho, Seongkuk
AU - Hwang, Eenjun
N1 - Funding Information:
ACKNOWLEDGMENT This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(No. NRF-2021R1A4A1031864)
Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - Convolutional neural networks have presented a new paradigm in the field of computer vision, but their fixed convolutional filter size causes long-range dependency loss problems. Also, because of the equal importance of all convolution filters, unnecessary information is included in the convolution result. These problems are more serious in image segmentation, where every pixel is labeled, as opposed to image classification or object recognition. In this paper, to alleviate these problems, we propose an attention-based neural network block that extracts robust features including channel and spatial information. The proposed block calculates the pixel relationship on a unit channel and then calculates the importance among channels. To evaluate the effectiveness of the proposed blocks, we constructed segmentation models by applying the proposed blocks to well-known network and measures their segmentation accuracy. We report some of the results.
AB - Convolutional neural networks have presented a new paradigm in the field of computer vision, but their fixed convolutional filter size causes long-range dependency loss problems. Also, because of the equal importance of all convolution filters, unnecessary information is included in the convolution result. These problems are more serious in image segmentation, where every pixel is labeled, as opposed to image classification or object recognition. In this paper, to alleviate these problems, we propose an attention-based neural network block that extracts robust features including channel and spatial information. The proposed block calculates the pixel relationship on a unit channel and then calculates the importance among channels. To evaluate the effectiveness of the proposed blocks, we constructed segmentation models by applying the proposed blocks to well-known network and measures their segmentation accuracy. We report some of the results.
KW - Attention mechanism
KW - Convolutional Neural Networks
KW - Feature extraction
KW - Image segmentation
UR - http://www.scopus.com/inward/record.url?scp=85124263263&partnerID=8YFLogxK
U2 - 10.1109/BigComp54360.2022.00054
DO - 10.1109/BigComp54360.2022.00054
M3 - Conference contribution
AN - SCOPUS:85124263263
T3 - Proceedings - 2022 IEEE International Conference on Big Data and Smart Computing, BigComp 2022
SP - 246
EP - 250
BT - Proceedings - 2022 IEEE International Conference on Big Data and Smart Computing, BigComp 2022
A2 - Unger, Herwig
A2 - Kim, Young-Kuk
A2 - Hwang, Eenjun
A2 - Cho, Sung-Bae
A2 - Pareigis, Stephan
A2 - Kyandoghere, Kyamakya
A2 - Ha, Young-Guk
A2 - Kim, Jinho
A2 - Morishima, Atsuyuki
A2 - Wagner, Christian
A2 - Kwon, Hyuk-Yoon
A2 - Moon, Yang-Sae
A2 - Leung, Carson
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2022 IEEE International Conference on Big Data and Smart Computing, BigComp 2022
Y2 - 17 January 2022 through 20 January 2022
ER -