TY - GEN
T1 - LR-SARNET
T2 - 18th Chinese Conference on Image and Graphics Technology and Application Conference, IGTA 2023
AU - Chang, Shibo
AU - Fu, Xiongjun
AU - Dong, Jian
AU - Chang, Hao
N1 - Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.
PY - 2023
Y1 - 2023
N2 - SAR Ship image has the characteristics of complex background, blurred target edge and scale difference, which makes the target detection difficult. This article builds a lightweight and robust network for multi-scale and multi-scene SAR ship detection (LR-SARNET). Firstly, with the goal to minimize the computational complexity of the model, a lightweight backbone feature extraction network (CGNet) is designed to generate sufficiently redundant feature maps with low computational cost. Secondly, a linear feature fusion module (ENECK) is designed to efficiently fuse deep local feature maps. Finally, the extremely efficient spatial pyramid (EESP) is integrated into the target detection head, which expands the receptive field of the network. The experiment on SSDD and HRSID dataset proves that our algorithm has strong robustness and excellent generalization performance.
AB - SAR Ship image has the characteristics of complex background, blurred target edge and scale difference, which makes the target detection difficult. This article builds a lightweight and robust network for multi-scale and multi-scene SAR ship detection (LR-SARNET). Firstly, with the goal to minimize the computational complexity of the model, a lightweight backbone feature extraction network (CGNet) is designed to generate sufficiently redundant feature maps with low computational cost. Secondly, a linear feature fusion module (ENECK) is designed to efficiently fuse deep local feature maps. Finally, the extremely efficient spatial pyramid (EESP) is integrated into the target detection head, which expands the receptive field of the network. The experiment on SSDD and HRSID dataset proves that our algorithm has strong robustness and excellent generalization performance.
KW - Anchor-free mechanism
KW - Intensive object detection
KW - SAR image
KW - Ship detection
KW - Small target detection
UR - http://www.scopus.com/inward/record.url?scp=85176015119&partnerID=8YFLogxK
U2 - 10.1007/978-981-99-7549-5_33
DO - 10.1007/978-981-99-7549-5_33
M3 - Conference contribution
AN - SCOPUS:85176015119
SN - 9789819975488
T3 - Communications in Computer and Information Science
SP - 456
EP - 471
BT - Image and Graphics Technologies and Applications - 18th Chinese Conference, IGTA 2023, Revised Selected Papers
A2 - Yongtian, Wang
A2 - Lifang, Wu
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 17 August 2023 through 19 August 2023
ER -