SRAOD: Super-Resolution-Based Object Detection in Multimodal Remote Sensing Imagery

This study explores the task of accurately detecting small objects in remote sensing imagery (RSI), which presents a significant challenge due to the need for robust feature extraction and the computational demands of complex deep neural networks. We introduce object detection methods designed to attain a speed and accuracy balance in RSI analysis. To address this issue, leverages a multimodal data fusion approach to extract complementary information from various data sources, enhancing its ability to detect small objects in RSI. The multimodal fusion (MF) process is symmetric and compact, ensuring efficient data integration significantly improves detection accuracy without introducing excessive computational overhead. Our approach minimizes the additional computational requirements, ensuring efficient, and speedy object detection can break through significant applications in fields such as remote sensing, where the accurate detection of small objects is crucial for various tasks. VEDAI RS dataset shows that SRAOD performs better than the most advanced models, such as YOLOv5l, YOLOv5x, and YOLOrs, in terms of accuracy. Furthermore, SRAOD achieves this accuracy while significantly reducing the model’s parameter size and computational demands, with an 18x reduction in parameter size and 3.8x fewer GFLOPs than YOLOv5x.