LLTQ plus : A Hardware-Friendly Quantization Framework for Modern YOLO Architectures

초록

YOLO-based object detection models are widely used in real-time applications due to their high accuracy and fast inference speed. However, their complex architectures and computational demands present challenges for deployment on low-power edge devices. To address this, we propose LLTQ+, an enhanced version of the hardware-friendly quantization method LLTQ. The proposed approach preserves batch normalization layers during Quantization-Aware Training (QAT) to maintain training stability and accuracy, and introduces a quantization strategy that preserves the representational capacity of RepConv, a key structural component of YOLO networks. Experimental results on the PASCAL VOC dataset demonstrate the effectiveness of LLTQ+. On YOLOv10-s, LLTQ+ achieved 80.6% mAP(0.5) and 61.8% mAP(0.5:0.95) under integer-only inference, surpassing LLTQ by 0.9 and 1.7 percentage points, respectively. On YOLOv9-t, LLTQ+ reached 52.9% mAP(0.5:0.95), an improvement of 0.5 points over LLTQ. Consistent performance gains were also observed across other architectures such as YOLOv7 and YOLOv7-tiny. These results confirm that LLTQ+ effectively supports integer quantization for even the latest and more complex YOLO networks, providing a practical quantization solution that balances accuracy and computational efficiency.

키워드