YOLO-TUF: An Improved YOLOv5 Model for Small Object Detection

Traditional target detection algorithms frequently encounter challenges in accurately detecting objects within complex, cluttered environments. This paper presents an optimized YOLOv5-based model to mitigate such limitations. Our contributions are threefold: Firstly, we enhance the upsampling procedure by amalgamating transposed convolution with the CBAM attention mechanism, fortifying the network’s fine-grained feature extraction capabilities. Secondly, we introduce an optimized feature-processing module, which enhances feature utilization while maintaining a lightweight architecture. Lastly, we integrate EfficientNet into the backbone architecture to amplify feature extraction performance. We validate our approach using the PASCAL VOC dataset, achieving an mAP0.5 of 84.00% and an mAP0.5:0.95 of 62.10%, while maintaining a modest parameter size of 13.22MB. These results mark an improvement of 4.50% ± 0.12% and 8.20% ± 0.09% over the benchmark, demonstrating an efficient trade-off between computational efficiency and detection accuracy. The proposed model outperforms conventional YOLOv5 algorithms and remains competitive with contemporary state-of-the-art object detection techniques. Code is available at https://github.com/chenxz0906chenxz/YOLO-TUF/ .