| 引用本文: | 王新良,王明旭,徐志宗.多尺度融合与感受野增强的生成人脸检测算法[J].哈尔滨工业大学学报,2026,58(9):93.DOI:10.11918/202409021 |
| Wang Xinliang,Wang Mingxu,Xu Zhizong.Generative face detection algorithm based on multi-scale fusion and receptive field enhancement[J].Journal of Harbin Institute of Technology,2026,58(9):93.DOI:10.11918/202409021 |
|
| 摘要: |
| 为减少滥用生成人脸技术导致的社交媒体欺诈行为及虚假新闻传播问题,本文提出基于YOLOv8的多尺度融合与感受野增强的生成人脸检测算法(MSERF-YOLOv8)。首先,引入逐通道卷积模块替换原网络中的卷积模块,利用其独立卷积核处理通道的特点,保持特征的独立性;其次,在主干网络高层结合扩张式残差模块构建全新的C2f模块,扩大主干网络高层感受野;再次,引入高效多尺度注意力机制,通过跨空间学习方法聚合多个并行子网络的输出特征图,丰富特征聚合;最后,基于ghost module、动态卷积和C2f模块构造新的混合架构,采用分组卷积参数共享的方式重构轻量化检测头。在5种GAN生成人脸自制数据集上开展实验,结果表明:MSERF-YOLOv8的均值平均精度达到了95.14%,相较于YOLOv8提升了1.25%,参数量仅为5.30×106,检测速度达到90.10帧/s;与经典目标检测算法和改进算法相比,该算法在特征提取和特征融合方面均有一定的优势。所提算法提高了生成人脸的检测精度、检测速度,同时节约了计算资源。 |
| 关键词: 生成人脸检测 YOLOv8 轻量化 特征提取 多尺度融合 注意力机制 |
| DOI:10.11918/202409021 |
| 分类号:TP391.4 |
| 文献标识码:A |
| 基金项目:河南省高等学校青年骨干教师培养计划(2019GGJS060);河南省高等学校重点科研基金(21B413005) |
|
| Generative face detection algorithm based on multi-scale fusion and receptive field enhancement |
|
Wang Xinliang1,Wang Mingxu1,Xu Zhizong1,2
|
|
(1.School of Physics & Electronic Information Engineering, Henan Polytechnic University, Jiaozuo 454000, Henan, China; 2.Power Supply Division, Hebi Coal Industry (Group) Co., Ltd., Hebi 458000, Henan, China)
|
| Abstract: |
| In order to reduce the social media fraud and false news dissemination caused by misuse of generative face technology, a generative face detection algorithm is proposed in this paper based on YOLOv8 with multi-scale fusion and receptive field enhancement(MSERF-YOLOv8). First, a depthwise convolution module is introduced to replace the convolution module in the original network, utilizing its independent convolution kernel to process the channels and maintain feature independence; second, a new C2f module is constructed by combining the expansive residual module at the higher level of the backbone network to expand the sensory field at the higher level of the backbone network; third, an efficient multi-scale attention mechanism is introduced to aggregate the output feature maps of multiple parallel sub-networks through a cross-space learning method to enrich the feature aggregation; finally, a new hybrid architecture is constructed based on the ghost module, dynamic convolution and C2f module, the grouped convolutional parameter sharing is utilized to reconstruct the lightweight detection head. Experiments were conducted on five self-made GAN-generated face datasets, and the results show that the mean average accuracy of MSERF-YOLOv8 reaches 95.14%, which is 1.25% higher compared to YOLOv8, with the number of references being only 5.30×106 and the detection speed 90.10 frames/s. Compared with the classical target detection algorithm and some improved algorithms, the proposed algorithm has certain advantages in both feature extraction and feature fusion. The proposed algorithm improves the detection accuracy and detection speed of generative faces, while saving computational resources. |
| Key words: generative face detection YOLOv8 lightweight feature extraction multi-scale fusion attention mechanism |