Lightweight Multi-Scale Feature Aggregation for Small Object Detection in Remote Sensing Images

Because of complex backgrounds, little effective information, and small target sizes in remote sensing images, recent object detection methods generally suffer from wrong detections of small objects. We propose a lightweight multi-scale feature aggregation algorithm for small object detection based on YOLOv8. The network structure of YOLOv8 is reconstructed by adding a small object detection branch and designing a multi-scale feature fusion module that employs grouped convolutions with different kernel sizes to extract features at different scales. A focal attention model is incorporated after the Spatial Pyramid Pooling (SPP) layer to adaptively gather contextual information from different focal levels. This lets the network focus more on small details about the structure. Experiments using the public VisDrone and AI-TOD datasets show that detection accuracy is improved by 5.4% and 4.0%. There was a 33% reduction in model parameters.