YOLO with attention mechanism for enhanced small object detection

Abstract

Multi-scale feature fusion and attention mechanisms are leveraged in this study to address the persistent challenge of small object detection in complex visual environments. To enhance the feature representation of small, low-contrast, and scale varying targets, the proposed detection framework is built on the YOLO framework and integrates a lightweight hybrid attention module and adaptive skip connections. By adopting selective channel-spatial attention and dynamic feature map alignment, the system retains key fine-grained details throughout the network, thus greatly reducing information loss during deep convolution processing. Experimental validation on the COCO small object subset and drone aerial images shows that the model achieves significant improvements in average precision and recall compared to established YOLO variants, especially for objects smaller than 32 x 32 pixels. The real-time inference speed and low memory usage are maintained, proving its effectiveness in embedded and resource-constrained applications. The effectiveness of combining advanced multi-scale fusion with attention mechanism to meet the needs of autonomous perception system is emphasized, and a robust and effective solution is provided for the scene where accurate small target detection is crucial.