ISOD: improved small object detection based on extended scale feature pyramid network

Rapid and accurate target detection is one of the inevitable requirements of intelligent construction site. To meet the speed requirements and improve detection accuracy, an improved small object detection (ISOD) network is proposed. The network utilizes an efficient channel attention mechanism to extract features in the backbone and combines the proposed extended scale feature pyramid network to simplify calculations and create additional high-resolution pyramid layers to improve the ability of detecting small targets. To verify the effectiveness of ISOD, experiments are conducted using the proposed Reflective Vest Scene Dataset and Tsinghua-Tencent 100K, achieving 0.425 and 0.635 mAP@0.5 \(-\) $-$ 0.95, respectively, exceeding the SOTA YOLOv7 model, demonstrating its excellent small target detection capability and scalability.