Small Object Detection by Synchronous Attention Swin Transformer with Channel Granularity Adaptive Mechanism

The essence of small object detection is to establish a mapping from the image pixels to the location and classification of targets. It is well known that a few valid pixels and complex backgrounds are the greatest challenges. This is because the intricate mapping cannot be formulated in a small object pixel space while facing a huge disturbance produced by background pixel spaces. To address these issues, this paper proposes a novel small object detection network named synchronous attention granularity Swin Transformer (SAG-ST). A synchronous attention ST block is proposed to elegantly integrate information from deep and shallow features. And the granularity adaptive ST block employs a channel granularity adaptive mechanism to mitigate background interference by adaptively applying self-attention with varying granularities for different channels. Finally, this paper creates a small object detection dataset based on unmanned aerial vehicles with different flight altitudes. The experiments are carried out on the created dataset and VisDrone dataset, and the experimental results show that our SAG-ST algorithm achieves the best detection accuracy.