Info-FPN: An Informative Feature Pyramid Network for object detection in remote sensing images
Silin Chen,Jiaqi Zhao,Yong Zhou,Hanzheng Wang,Rui Yao,Lixu Zhang,Yong Xue
DOI: https://doi.org/10.1016/j.eswa.2022.119132
IF: 8.5
2022-11-07
Expert Systems with Applications
Abstract:Feature pyramid networks are widely applied in remote sensing images for object detection to deal with the challenge of large scale variation in objects. However, the feature pyramid-based object detector for remote sensing images ignores the channel information loss, feature misalignment, and additional computational overhead to eliminate the aliasing effect, leading to inadequate feature extraction for multi-scale objects in remote sensing images. To address these challenges, an Informative Feature Pyramid Network (Info-FPN) is proposed. Specifically, we propose a Pixel Shuffle-based lateral connection Module (PSM) for the complete preservation of channel information in the feature pyramid. Then, to alleviate the problem of confusion caused by feature misalignment, a Feature Alignment Module (FAM) is proposed to achieve aligned feature fusion by template matching and learning feature offsets in the feature fusion stage. To eliminate the aliasing effect, we design a Semantic Encoder Module (SEM), which reduces the parameters and computation of model with the desirable detection accuracy. Extensive experiments on two challenging remote sensing datasets, namely DOTA and HRSC2016, prove the effectiveness of the proposed method which achieves comparable detection performance with the state-of-the-art FPN-based method.
computer science, artificial intelligence,engineering, electrical & electronic,operations research & management science