| 李绪涛,邓耀华.电池字符缺陷检测DDP-YOLOv8模型方法[J].电子测量与仪器学报,2025,39(6):165-173 |
| 电池字符缺陷检测DDP-YOLOv8模型方法 |
| DDP-YOLOv8 model for battery character defect detection |
| |
| DOI： |
| 中文关键词: 字符缺陷 YOLOv8 注意力 动态采样 动态检测头 |
| 英文关键词:character defect YOLOv8 attention dynamic sampling dynamichead |
| 基金项目:国家自然科学基金(52175457)、广东省基础与应用基础研究基金(2022B151520053)、东莞市重点领域研发项目（20221200300042）资助 |
|
|
|
|
| 摘要点击次数: 10 |
| 全文下载次数: 5 |
| 中文摘要: |
| 针对消费电池产品表面字符缺陷检测中存在的缺陷位置动态分布、多尺度适应性差和细小缺陷识别困难等关键技术难题,提出了一种创新性的
可变形大核卷积注意力、动态采样和P2-动态检测头的YOLOv8（DDP-YOLOv8）检测模型框架。首先,针对YOLOv8在特征提取过程中无法有效调整特征图权重的问题,设计DCNv3-LKA注意力模块,通过融合动态卷积网络与大核注意力机制,在特征提取阶段实现空间权重自适应调整。其次,针对YOLOv8颈部网络在字符缺陷检测中的采样位置固定和多尺度适应性差的问题,对YOLOv8的颈部网络结构进行重构,采用跨尺度特征融合模块（CCFM）架构并提出一种引入了动态偏移量与可学习采样权重双驱动机制的动态采样器DS(DS-CCFM模块),突破传统特征金字塔的固定几何约束。最后,针对消费电池产品表面字符小尺度及YOLOv8检测头使用普通卷积层导致的特征表达不足与信息丢失问题,增加P2小目标检测层并在检测头融入DynamicHead多个自注意力机制(P2-DynamicHead模块),提升对微小缺陷的捕获能力。实验结果表明,DCNv3-LKA、DS-CCFM和P2-DynamicHead模块分别使模型在字符缺陷数据集上的平均精度均值(mAP）mAP@0.5达到91.8%、91.2%和92.4%,相较于YOLOv8n分别提高了1.7%、1.1%和2.3%。DDP-YOLOv8最终实现了94.0%的mAP@0.5,相较于基准模型YOLOv8n提升了3.9%,模型检测速度为85.1 fps,满足电池大规模定制生产中字符缺陷检测对高精度与实时性的需求。 |
| 英文摘要: |
| To address the critical challenges in surface character defect detection of consumer batteries, including dynamic defect localization, multi-scale adaptability, and fine-scale defect recognition, this paper proposes an innovative DDP-YOLOv8 framework. Firstly, to resolve the limitation of YOLOv8 in effectively adjusting feature map weights during feature extraction, we design a DCNv3-LKA attention module to achieve adaptive spatial weight adjustment through dynamic convolution and large-kernel attention fusion. Secondly, aiming to overcome the fixed sampling positions and poor multi-scale adaptability of YOLOv8’s neck network in character defect detection, we restructure the neck architecture by adopting a CCFM framework and propose a dynamic sampler (DS-CCFM module) incorporating dual-driven dynamic sampling mechanism. Finally, to mitigate the insufficient feature representation and information loss caused by standard convolution layers in YOLOv8’s detection head when handling small-scale battery characters, we introduce a P2 small-target detection layer and integrate multiple self-attention mechanisms from DynamicHead into the detection head (P2-DynamicHead module) to improves small defect recognition. Experimental results demonstrate that the DCNv3-LKA, DS-CCFM, and P2-DynamicHead modules achieve mean average precision (mAP) mAP@0.5 of 91.8%, 91.2%, and 92.4% respectively on the character defect dataset, representing improvements of 1.7%, 1.1%, and 2.3% over baseline YOLOv8n. DDP-YOLOv8 achieves a final mAP@0.5 of 94.0%, representing a 3.9% improvement over the baseline model YOLOv8n. With an FPS of 85.1, the model meets the requirements of high accuracy and real-time performance for character defect detection in large-scale customized battery production. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
