Employing deep learning for automatic river bridge detection from SAR images based on Adaptively effective feature fusion

International Journal of Applied Earth Observations and Geoinformation

卷:102

Ru Luo¹  Zhihui Yuan²  Ting Weng³  Zhenhong Li³  Lifu Chen⁴  Jin Xing⁴  Zhouhao Pan⁵  Siyu Tan⁵ 

[1] College of Geological Engineering and Geomatics, Chang’an University, Xi’an, China;

[2] Corresponding author at: School of Electrical and Information Engineering, Room B308, Engineering Building 1, 960 Wanjiali South Road, Changsha, China.;

[3] Technology, Changsha, China;

[4] School of Electrical and Information Engineering, Changsha University of Science &

[5] School of Engineering, Newcastle University, Newcastle upon Tyne NE1 7RU, UK;

关键词: River bridge detection;    Deep learning;    SAR image analytics;    Feature fusion;    Attention mechanism;   

DOI  :  

来源: DOAJ

【 摘 要 】

Automatic river bridge detection is a typical and valuable application for SAR image analysis. However, the background information of SAR image is complex, and there are many specious targets with similar features, such as road, ponds and ridges, which usually cause false alarms. And current river bridge detection methods fail to handle these interference efficiently. Therefore, this paper applies deep learning to SAR and proposes a new river bridge detection algorithm, which is named as Single Short Detection-Adaptively Effective Feature Fusion (SSD-AEFF). It can effectively reduce the interference of noisy information, and achieve fast and high-precision detection of river bridges in complex SAR imagery. SSD-AEFF is based on SSD, and AEFF module is innovated to enhance the multi-scale feature maps together with effective Squeeze-Excitation (eSE) module to further fuse effective features and decrease the interference of background information. Additionally, Non-Maximum Suppression (NMS) is used to screen out redundant candidate boxes to produce the final detection result. Moreover, Gradient Harmonizing Mechanism (GHM) loss function is introduced to solve the problem of sample imbalance in the training process. Experimental results on TerraSAR data compared with existing baseline models demonstrate the superiority of the proposed SSD-AEFF algorithm.

【 授权许可】

Unknown