Small Object Detection Network with Spatial Compression and Channel Enhancement Fusion

To overcome the low detection accuracy of small objects in existing remote-sensing target-detection algorithms, which leads to missed and false detections, an improved aerial small target detection algorithm based on YOLOv8n is proposed. This algorithm was enhanced using the SAHI slicing auxiliary inference technique to enhance the detection of small targets. By optimizing the network backbone to enhance the high-resolution feature-extraction and refine the neck structure, the algorithm enhances the perception of small targets. To emphasize the target features while mitigating the impact of background interference, the convolutional module was optimized using receptive-field attention and convolutional block attention module mechanisms, which highlight the spatial characteristics of the receptive field to strengthen the learning capability for small target features. The C2f was reengineered with a space-to-depth module to create a new feature reorganization and enhancement module, reducing the loss of fine-grained information and effectively aggregating local feature information. Experiments on the VisDrone2019 dataset show that the improved model increased the mAP₅₀ by 9.2% compared to the original YOLOv8n and outperformed other mainstream networks in terms of accuracy. Generalization experiments were conducted using the Tinyperson and Widerperson datasets. The results indicate that the optimized model significantly improved the detection precision for small targets compared to the original model.