Superyolo: Super Resolution Assisted Object Detection in Multimodal Remote Sensing Imagery

Abstract

Finding small things quickly and correctly in remote sensing pictures (RSI) is very hard because you need to use strong feature extraction and complex deep neural networks need a lot of computing power. The study introduces SuperYOLO, a novel approach to identifying objects that seeks to achieve a good combination of speed and accuracy in RSI analysis. SuperYOLO uses a multimodal data fusion method to combine information from different data sources in a way that makes it better at finding small items in RSI. This multimodal fusion (MF) process is both symmetric and compact, which makes it easy to combine data. SuperYOLO has an enhanced super-resolution (SR) learning branch in addition to MF. This SR branch lets the model make high-resolution (HR) feature representations, which lets it tell small items apart from the background when the input is low-resolution (LR). This makes recognition much more accurate without adding too much work to the computer. One great thing about SuperYOLO is that the SR branch is only used during training and is thrown away during inference. This method reduces the need for extra computing power, making sure that object recognition works quickly and efficiently. When tested on the well-known VEDAI RS dataset, SuperYOLO does better at accuracy than cutting-edge models like YOLOv5l, YOLOv5x, and YOLOrs. Additionally, SuperYOLO gets this level of accuracy while greatly lowering the model's parameter size and processing needs. Compared to YOLOv5x, SuperYOLO has 18 times fewer parameters and 3.8 times fewer GFLOPs. To sum up, SuperYOLO makes a strong case for choosing between accuracy and speed when it comes to finding small objects in RSI. The model does a better job than other options because it combines multimodal data fusion with assisted SR learning in a way that makes it more efficient and less complicated to use. This big step forward could have big effects in areas like remote sensing, where finding small things accurately is important for many jobs.