Enhancing Multi-Class Object Recognition with MultiCA-YOLOv7: A Comprehensive Study
Nan Wang · 2023
Computer vision and object detection techniques have achieved significant success across various domains. However, challenges posed by multi-class and complex multi-object scenarios often remain overlooked in model predictions. The diversity of target categories, coupled with factors like lighting variations, varying angles, and mutual occlusions, present challenges for achieving high-level recognition accuracy in existing methods. Addressing these challenges, this paper introduces MultiCA-YOLOv7, based on a multi-point embedded Coordinate Attention (CA) mechanism. The powerful feature extraction capability of the network backbone and the dual-channel and spatial attention enhancement of the CA module result in more accurate features. In this framework, MultiCA-YOLOv7 achieves an impressive 96.15% mAP50 on the test set without introducing additional parameters, outperforming the best-performing baseline model by 6.85%. Across six different categories in the test dataset, all models undergo rigorous evaluation. The experimental results indicate that the proposed model outperforms baseline models in most evaluation metrics. We analyzed the reasons for the ineffective performance of the CA attention mechanism based on YOLOv8. Furthermore, in the prediction of multi-level object detection, the comprehensively trained model successfully captures all moving targets without omission, validating the proposed model's higher accuracy and robustness.