| 钟锐,刘林彬,陈斌.基于特征拟合蒸馏度量网络的低分辨率人脸识别[J].高技术通讯(中文),2026,36(7):688~700 |
| 基于特征拟合蒸馏度量网络的低分辨率人脸识别 |
| Feature fitting distillation metric network for low-resolution face recognition |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 07. 003 |
| 中文关键词: 特征拟合蒸馏； 相似度量蒸馏； 低分辨率人脸识别； 非限制性场景 |
| 英文关键词: feature fitting distillation, similarity metric distillation, low-resolution face recognition, non-restrictive scenarios |
| 基金项目: |
| 作者 | 单位 | | 钟锐 | (赣南师范大学数学与计算机科学学院赣州 341000) | | 刘林彬 | | | 陈斌 | |
|
| 摘要点击次数: 64 |
| 全文下载次数: 41 |
| 中文摘要: |
| 在实际应用场景中，人脸识别系统很容易采集到低分辨率的人脸图像，这些图像中的面部细节特征大量丢失，同时还存在光照、遮挡以及姿态等干扰因素的叠加，将导致现有经典的人脸识别算法难以取得满足实际应用需求的识别精度。针对上述问题，提出了一种特征拟合蒸馏度量(feature fitting distillation metric，FFDM)网络。首先，在教师网络与学生网络的多个阶段的卷积层之间进行特征拟合蒸馏，使学生网络各阶段卷积层所提取特征尽可能接近教师网络各阶段卷积层特征，从而达到增强学生网络对低分辨率样本的特征描述能力的目标；随后，在教师网络和学生网络的相似度量子网间引入相似度量蒸馏损失函数，使学生网络通过该蒸馏损失函数学习到教师网络强大的细粒度特征的分类能力，使学生网络能够在非限制性复杂场景中具备较强的鲁棒性。最后，在多个数据集上的实验结果表明，所提出的 FFDM模型在低分辨率人脸识别任务中展现出更高的识别精度和泛化能力。 |
| 英文摘要: |
| Face recognition systems often capture low resolution facial images in practical applications where a significant amount of facial detail features are lost. Additionally, the presence of superimposed interference factors, such as illumination variations, occlusions, and pose changes, makes it challenging for existing classical face recognition algorithms to achieve satisfactory recognition accuracy that meets real-world application requirements. To address these issues, we propose a feature fitting distillation metric (FFDM) network. First, feature fitting distillation is performed between multiple stages of convolutional layers in both the teacher network and student network, enabling the features extracted by each convolutional layer in the student network to approximate those of the corresponding layers in the teacher network as closely as possible. This approach aims to enhance the student network’s capability in characterizing features from low-resolution samples. Subsequently, a similarity metric distillation loss function is introduced between the similarity measurement sub-networks of the teacher and student networks. The student network learns the powerful classification capability of the teacher network for fine-grained features through this distillation loss function, thus enabling the student network to have high robustness in complex and unconstrained scenarios. Finally, the experimental results on multiple datasets demonstrate that FFDM model exhibits higher accuracy and superior generalization capability in low-resolution face recognition tasks. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
