A lightweight channel bias for small object detection in aerial images

Small-object detection in drone-based imagery is difficult because small targets occupy very few pixels and often appear in cluttered backgrounds. Existing approaches often rely heavily on spatial cues, but their effectiveness is limited when the spatial information itself is unstable or insufficient. This paper introduces a simple Channel Bias (CB) mechanism that models relationships along the channel dimension to compensate for weak spatial cues. The method uses global average pooling and a lightweight 1 × N convolution to generate a channel-wise bias, which is broadcast and combined with the original features to highlight useful channels. Experiments on the VisDrone benchmark show that integrating Channel Bias mechanism into YOLO-based detectors yields consistent accuracy gains. These results indicate that channel-level information is a practical way to improve small-object detection when spatial evidence is limited.