MSA-YOLOv5: An Improved Light-Weight YOLOv5 Model for Small Object Detection

Abstract

Recent object detection research has primarily focused on achieving high accuracy, often overlooking model efficiency and size. To address this issue, we propose the Multi-SpatialLite Attention (MSA) YOLOv5 model, which emphasizes a balance between accuracy and model size. The proposed model was trained on the VHR-10 dataset for small object detection, achieving a lightweight design while maintaining excellent performance in small object detection. Experimental results show that the proposed model achieved 97.6% precision, 99.0% recall, and 97.0% F1-score, while maintaining a lightweight structure with 1.795 million parameters and demonstrating outstanding performance with a mean Average Precision (mAP) of 98.3%. These results suggest the model's ability to enhance precise detection of small objects and provide effective real-time performance even in limited hardware environments. Notably, it has surpassed existing models in applications requiring real-time performance in constrained hardware environments. © 2024 IEEE.