基於 ReActNet 主幹 YOLOv8 硬體加速設計;YOLOv8 hardware acceleration design based on ReActNet backbone

NCUIR > College of Electrical Engineering & Computer Science > Executive Master of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/98160

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98160

Title:	基於 ReActNet 主幹 YOLOv8 硬體加速設計;YOLOv8 hardware acceleration design based on ReActNet backbone
Authors:	陳柏家;Chen, Bo-Jia
Contributors:	資訊工程學系在職專班
Keywords:	物件偵測;二值卷積;剪枝;量化;硬體加速
Date:	2025-06-16
Issue Date:	2025-10-17 12:26:56 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著深度學習在目標檢測領域的快速發展，YOLOv8 憑藉優異的辨識準確度被廣泛應用，然而其龐大的參數量限制了推論效率。雖然常見壓縮技術如二值化可有效縮減模型體積並提升速度，卻經常伴隨精度下降，難以兼顧效能與準確性。有鑑於此，本研究提出輕量化模型 RYv8_BNH，結合 ReActNet 二值化策略、剪枝與量化技術，並於檢測頭中設計輕量卷積模組 RYv8_BNH_LConv，以在壓縮參數的同時維持辨識性能。更導入 MIAT 方法論進行系統架構設計，透過模組化與管線化方式實現硬體加速。實驗結果顯示，RYv8_BNH 模型在 Pascal VOC 資料集上的 mAP50-95 為 39%，參數量僅為 0.46M。相較於其他輕量模型，準確度提升約 4%，參數量最多減少達 86%。硬體加速後的推論時間為 10.08 μs 較 PyTorch 環境的 3400 μs 大幅提升。整體實驗結果驗證本模型在推論速度與辨識精度間達成良好平衡，展現其於即時性與辨識精度要求的資源受限場景之發展潛力。;With the rapid advancement of deep learning in the field of object detection, YOLOv8 has been widely adopted due to its outstanding recognition accuracy. However, its large number of parameters limits inference efficiency, posing challenges for real-time applications. Although common compression techniques such as binarization can effectively reduce model size and improve speed, they often lead to a decline in accuracy, making it difficult to balance efficiency and precision. To address this issue, this study proposes a lightweight model, RYv8_BNH, which integrates ReActNet-based binarization strategies along with pruning and quantization techniques. A lightweight convolutional module, RYv8_BNH_LConv, is designed in the detection head to preserve recognition performance while reducing parameter complexity. The MIAT methodology is also introduced to guide the system-level architecture design, enabling hardware acceleration through modular and pipelined implementation. Experimental results show that the RYv8_BNH model achieves 39% mAP50–95 on the Pascal VOC dataset, with only 0.46 million parameters. Compared to other lightweight models, it achieves approximately a 4% improvement in accuracy while reducing parameter count by up to 86%. The hardwareaccelerated inference time is 10.08 μs, a significant improvement over the 3400 μs observed in the PyTorch environment. These results confirm that the proposed model achieves a strong balance between inference speed and detection accuracy, demonstrating its potential for realtime, resource-constrained applications requiring both efficiency and precision
Appears in Collections:	[Executive Master of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	3	View/Open

社群 sharing

Loading...