Spatial-Transformer and Cross-Scale Fusion Network (STCS-Net) for Small Object Detection in Remote Sensing Images

• Published in: Journal of the Indian Society of Remote Sensing 2023-07, Vol.51 (7), p.1427-1439
• Main Authors: Lan, Jinhui, Zhang, Cheng, Lu, Weijian, Gu, Naiwei
• Format: Article
• Language: English
• Subjects: Accuracy; Earth and Environmental Science; Earth Sciences; Feature extraction; Machine learning; Modules; Object recognition; Remote sensing; Remote Sensing/Photogrammetry; Research Article; Transformers
• ISSN: 0255-660X; 0974-3006
• DOI: 10.1007/s12524-023-01709-w

With the rapid development of deep learning in recent years, the detection technology of objects in natural-scene images has been greatly improved. However, the scale of remote sensing image objects is diverse and many small objects exist in the images, which results in the low overall accuracy of remote sensing image object detection and omission of small objects. To address this issue, we propose a novel method for small object detection in optical remote sensing images. In this method, a spatial-transformer module is constructed using spatial attention and self-attention to realize feature extraction of key regions in the image space. Then, a cross-scale feature-fusion module is constructed in the neck region of the network to fuse similar feature information at different levels. Finally, the optimal bounding box is selected by reconstructing the bounding box screening rules and ignoring the interference of redundant bounding boxes. We experimentally evaluate our method on the remote sensing image dataset DIOR, DOTA and AI-TOD, and we compare it with other state-of-the-art object detection methods. In addition to improving the overall accuracy by 2%, 2.9% and 2.7% respectively, the proposed method can also considerably improve the small object detection accuracy.