Spatial attention and wide activation-based deep super-resolution scheme for vehicle detection in noisy and low-resolution LEO satellite imagery

Abstract

This paper presents a robust vehicle detection method for Low Earth Orbit satellite imagery degraded by environmental factors. To enhance feature representation from RGB and IR images, a denoising autoencoder and a multimodal fusion (MF) module with multi-head squeeze excitation are used. To compensate for texture loss in small vehicles at low resolution, a spatial attention module (SAM) is embedded into the YOLOv5 backbone, and a modified wide activation super-resolution (WDSR) branch is attached for learning finer-grained representations. Experiments on degraded images show that the proposed model outperforms the YOLO series and SuperYOLO in terms of mean average precision (mAP).