Integrating Transformer and Enhanced YOLOv10 for Improved Small Object Detection in RGB and Infrared Remote Sensing Imagery

Small object detection in remote sensing imagery is crucial for applications such as environmental monitoring and disaster response but is challenged by small object size, low contrast, occlusions, and varying lighting conditions. To address these issues, we integrate the strengths of SuperYOLO in super-resolution and multimodal data fusion with the advanced features of YOLOv10. Additionally, we have modified YOLOv10's head to only use the P3 layer, a specialized layer for small object detection, instead of using all P3, P4, and P5 layers. This improves the model's ability to capture and identify small objects while reducing the number of parameters. Specifically, we have added the transformer block (C3TR) to our proposed model to enhance task accuracy. Experimental evaluations on the widely used VEDAI dataset confirm the effectiveness of our approach. Our proposed model achieves an accuracy of 85.46% (in terms of mAP50), surpassing state-of-the-art variants like YOLOv8 and SuperYOLO by over 3%. Additionally, our model's parameter size and GFLOPs are approximately 8.23% and 13.2% less, respectively, compared to the YOLOv10 backbone.