| 基于轻量化和特征融合的隐私信息识别 |
 点此下载全文 |
| 引用本文：李超良，翁昌豪，吴林远，胡贤燊.基于轻量化和特征融合的隐私信息识别[J].计算技术与自动化,2026,(2):119-125 |
| 摘要点击次数: 42 |
| 全文下载次数: 0 |
|
|
| 中文摘要:当前社交平台图片存在隐私泄露风险，而现有目标检测模型存在参数量大、信息提取效率低和多尺度特征融合不足等问题。为此，提出了改进的GEA-YOLOv8n模型。首先，引入幽灵卷积优化骨干网络，通过线性变换处理冗余信息，降低计算量和参数量。其次，设计EC2F机制，结合CSPDarknet53与两阶段FPN结构，在保持通道维度的同时增强跨通道信息交互能力。最后，采用自适应空间特征融合策略，动态融合多尺度特征。实验表明，改进模型在精度和召回率上均优于基准模型YOLOv8n，能有效检测社交网络中的隐私内容，为自动化审核提供技术支撑。 |
| 中文关键词:目标识别；YOLOv8 注意力机制；轻量化算法 |
| |
| Privacy Information Recognition Based on Lightweighting and Feature Fusion |
|
|
| Abstract:The current social media platforms’ images are at risk of privacy leakage, and the existing object detection models have problems such as large parameter quantity, low information extraction efficiency, and insufficient multi-scale feature fusion. Therefore, this study proposes an improved GEA-YOLOv8n model. Firstly, the ghost convolution is introduced to optimize the backbone network, which processes redundant information through linear transformation, reducing the computational load and parameter quantity. Secondly, the EC2F (Efficient Channel Attention - CSPDarknet53 to two-Stage FPN) mechanism is designed, combining CSPDarknet53 with the two-stage FPN structure, enhancing the cross-channel information interaction ability while maintaining the channel dimension. Finally, the adaptively spatial feature fusion (ASFF) strategy is adopted to dynamically fuse multi-scale features. Experiments show that the improved model outperforms the baseline model YOLOv8n in terms of precision and recall rate, effectively detecting privacy content in social networks and providing technical support for automated review. |
| keywords:target recognition YOLOv8 attention mechanism lightweighting algorithm |
| 查看全文 查看/发表评论 下载pdf阅读器 |
|
|
|
