Abstract:

Images in real-world scenarios are affected easily by external lighting conditions or camera parameters， resulting in issues such as low overall brightness， poor visual effects， and much noise. These problems lead to difficulties in subsequent face recognition tasks， thereby causing engineering challenges. Therefore， an adaptive low-light face recognition network based on feature fusion， named LLANet （Low Light Adaptive face recognition Network）， was proposed with four parts： a decomposition subnet， a restoration subnet， an adjustment subnet， and a backbone network. Low-light and normal-light images were used as inputs. Firstly， based on Retinex theory， the input low-light and normal-light images were decomposed into the corresponding illumination and reflection maps. The illumination map was input into the adjustment subnet， where an attention mechanism was introduced to focus on lighting features， thereby enhancing the performance of low-light image enhancement and ensuring quality of the enhanced images. At the same time， the reflection map was input into the restoration subnet for detail restoration and noise reduction， thereby addressing degradation and noise issues of the reflection map in low-light images. And features of output of the adjustment and restoration subnets were fused to obtain the enhanced feature map. Then， to accomplish downstream face recognition tasks as well as prevent overfitting of lighting features and inaccuracies in face feature extraction， a weighted feature fusion strategy was adopted to combine the original face features extracted by the backbone network with the enhanced feature map， resulting in a feature map with richer information. Finally， an Adversarial Data Augmentation （ADA） strategy was introduced to generate more hard samples during training， thereby addressing the ill-posed problem while reducing the influence of alignment errors caused by low-light images during face detection phase， as a result， the network performance was further improved. Experimental results on CASIA-FaceV5， SoF， and YaleB low-light face datasets demonstrate that LLANet has the recognition rates reached 94.67%， 98.22%， and 97.24%， respectively， which are improved by 2.14， 1.58， and 2.10 percentage points on the three datasets， respectively， compared with ARoFace （Alignment Robust Face）. It can be seen that LLANet achieves high recognition accuracy in low-light scenarios.

Key words: low-light image enhancement, face recognition, attention mechanism, feature fusion, Adversarial Data Augmentation (ADA)

摘要：

现实场景中图像容易受外部光线条件或相机参数的影响而出现图像整体亮度过低、视觉效果不好和噪声多等问题，导致后续的人脸识别任务出现困难，从而引发工程问题。为此，针对低光照场景下的人脸识别任务，提出一种基于特征融合的低光照场景下的自适应人脸识别网络LLANet（Low Light Adaptive Face Recognition Network），该网络由分解子网络、恢复子网络、调节子网络和主干网络这4个部分组成。首先，根据Retinex理论将输入的低光照图像和正常光照图像分解为对应的照度图与反射图：照度图被输入调节子网络，在调节子网络中引入注意力机制，使网络聚焦于光照特征，不仅能提升低光照图像增强性能，还能确保增强图像的质量；反射图则输入恢复子网络，进行细节恢复与降噪操作，有效解决低光照图像反射图退化和噪声问题，将调节子网络与恢复子网络的输出进行特征融合，得到增强后的特征图。其次，为完成下游识别任务，且防止光照特征过拟合和人脸特征提取不准确，采用加权特征融合策略，融合主干网络提取的原始人脸特征与增强后的特征图，获得信息更丰富的特征图。最后，引入对抗性数据增强（ADA）策略，在训练时生成更多困难样本，在解决不适定问题的同时降低低光照图像在人脸检测阶段的对齐误差对网络的影响，进一步提升网络性能。在CASIA-FaceV5、SoF和YaleB这3个低光照人脸数据集上的实验结果表明，LLANet的识别率分别达到了94.67%、98.22%和97.24%，与ARoFace（Alignment Robust Face）相比，分别提高了2.14、1.58和2.10个百分点。可见，LLANet在低光照场景下有着较高的识别率。

关键词: 低光照图像增强, 人脸识别, 注意力机制, 特征融合, 对抗性数据增强