Authors - Nhat Minh Nguyen, Tu Anh Nguyen, Huy Thai Hinh, Quang Nhat Nguyen, Thong Minh Phuc Nguyen, Vinh Dinh Nguyen
Abstract - Car logo detection is crucial for intelligent transportation systems. This paper compares attention mechanisms integrated into YOLOv8m for car logo detection, evaluating Efficient Channel Attention (ECA), SimAM, and their combination against a baseline YOLOv8m (95.2%
[email protected], 48.6%
[email protected]:0.95, 79.3 GFLOPs). On a dataset of 5372 images, ECA achieves 94.5%
[email protected] and 50.2%
[email protected]:0.95 with 79.1 GFLOPs, SimAM reaches 96.1%
[email protected] and 50.0%
[email protected]:0.95 with 79.1 GFLOPs, and the combined approach attains 95.3%
[email protected] and 50.1%
[email protected]:0.95 with 79.3 GFLOPs. All variants maintain real-time performance with inference speeds of 29.29 FPS (ECA), 28.69 FPS (SimAM), and 27.95 FPS (combined), compared to the baseline’s 28.32 FPS on an NVIDIA RTX 3050 laptop GPU. These results demonstrate that integrating lightweight attention mechanisms can substantially enhance detection performance in data-constrained scenarios while preserving real-time efficiency.
