Based on the Optimization and Performance Evaluation of YOLOv8 Object Detection Model with Multi-backbone Network Fusion
Jisong Ye, Yanjuan Wu, Rong Wang
Abstract
The paper aims to enhance the recognition accuracy of the YOLOv8 object detection model in complex scenarios. To achieve this goal, various high-performance backbone networks, including ResNet50, DenseNet169, ConvNeXt, and EfficientNetv2, are integrated with YOLOv8 to construct four novel detection models: YOLOv8-ResNet50, YOLOv8-DenseNet169, YOLOv8-ConvNeXt, and YOLOv8-EfficientNetv2. These models combine the unique characteristics of each backbone network, aiming to further improve detection accuracy while maintaining YOLOv8's real-time performance. Rigorous experimental validation is conducted on a self-constructed leaf mustard dataset. The experimental results demonstrate that YOLOv8-EfficientNetv2 performs the best among these models, achieving a high accuracy of 95.2% in mAP50 and 85.3% in mAP50:95. Compared with the original YOLOv8, YOLOv8-EfficientNetv2 exhibits improvements of 0.87% and 1.6% in mAP50 and mAP50:95, respectively, significantly enhancing the accuracy of object detection. This research provides novel ideas and methods for the application of YOLO series models in complex scenarios, laying a solid foundation for future object detection research.