Research on small target detection models based on YOLO

Target detection technology has achieved significant progress, yet the detection of small target remains a challenging research area. To address the difficulty of detecting small target by models, an improved model for small target detection based on YOLO series was proposed. A pre-classification strategy was introduced to reduce interference between feature layers, and a Coord Attention module was embedded in the network to improve the performance of feature extraction. Additionally, a global residual structure was incorporated to ensure the integrity of the original image features. The model was tested on the publicly available VOC2007+2012 dataset, and the findings revealed that: ① the overall test set images of the improved YOLOv5 and YOLOv7 network models achieved average precision improvements of 4.5% and 0.8%, respectively, compared to the original network models; ② the detection accuracy for small targets was increased by 3.4% and 6.3%, respectively, outperforming the original network models. The results indicate that the proposed model exhibits robustness and accuracy in target detection tasks.