<![CDATA[Deep learning methods in computer vision play a crucial role in object localization using camera-based sensors, with Convolutional Neural Networks serving as the dominant approach for object detection. However, many existing models incur high computational costs due to deep architectures and complex operations, limiting their use for real-time deployment on low-cost, resource-constrained devices. The YOLOv12 architecture offers lightweight variants to improve computational efficiency. This study evaluates the trade-off between efficiency and detection performance by comparing model variants using the number of parameters, floating-point operations, and inference speed, while detection accuracy is measured using mean average precision. The results assess the suitability of lightweight models for real-time deployment in resource-constrained environments such as underwater monitoring and conservation. Experimental results on the Real-World Underwater Object Detection dataset demonstrate that YOLOv12-nano achieves 5.7% lower accuracy compared to YOLOv12-medium but requires only 2.57 million parameters and 6.5 GFLOPs, significantly less than YOLOv12-medium with 20.1 million parameters and 67.8 GFLOPs. Moreover, YOLOv12-small requires 9.26 million parameters and 21.5 GFLOPs, positioning it between nano and medium in terms of complexity while still maintaining competitive accuracy. In the inference process, YOLOv12-nano achieves 16.48 FPS on a 12th Gen Intel(R) Core (TM) i5-12450HX CPU. In comparison, YOLOv12-small runs at 6.28 FPS, while YOLOv12-medium runs at 2.36 FPS. These results indicate that YOLOv12-nano is the most suitable variant for real-time deployment on CPU-based platforms.]]>
Copyrights © 2025
