摘要
针对现有人脸识别模型无法从戴口罩人脸中有效提取区域特征问题,提出融合双重注意力机制的戴口罩人脸识别模型。首先将自建的戴口罩人脸图像作为输入数据,以ResNet50为基准网络,向残差块中引入协调注意力与分割注意力机制。其中协调注意力用于减少口罩区域特征提取,降低口罩区域特征干扰;分割注意力用于细粒度提取非口罩区域特征,从关键部位提取更多特征。然后使用ArcFace分类函数优化分类边界,再结合交叉熵损失函数作为约束,实现戴口罩人脸精细识别。实验结果表明,本文模型在测试集取得95.2%的识别准确率,与ResNet50、AttentionNet模型相比,识别准确率分别提高1个百分点、1.5个百分点。
To address the problem that existing face recognition models cannot effectively extract regional features from faces wearing masks,a face recognition model incorporating a dual attention mechanism is proposed for faces wearing masks.Firstly,a self-constructed face image wearing a mask is used as input data,and ResNet50 is used as the benchmark network to introduce coordinate attention and split attention mechanisms into the residual blocks,where coordinate attention is used to reduce feature extraction in the mask region and reduce feature interference in the mask region;Split attention is used to extract non-mask re⁃gion features at a fine granularity and extract more features from key areas.The ArcFace classification function is then used to op⁃timize the classification boundary,combined with a cross-entropy loss function as a constraint,to achieve fine-grained recogni⁃tion of faces wearing masks.The experimental results show that the model in this paper achieves 95.2%recognition accuracy in the test set,which is 1 percent point and 1.5 percent point higher than that of ResNet50 and AttentionNet models respectively.
作者
盛江岸
陈淑荣
SHENG Jiang-an;CHEN Shu-rong(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)
出处
《计算机与现代化》
2023年第2期72-77,共6页
Computer and Modernization
基金
陕西省重点研发计划项目(2022GY-039)。
