Abstract: Images in real-world scenarios are easily affected by external lighting conditions or camera parameters, resulting in issues such as low overall brightness, poor visual effects, and high noise levels. These problems lead to difficulties in subsequent face recognition tasks, causing engineering challenges. To address this, a low-light adaptive face recognition network based on feature fusion, named LLANet (Low Light Adaptive face recognition Network), was proposed for face recognition tasks in low-light scenarios. LLANet was composed of 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. First, based on Retinex theory, the images were decomposed into 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, enhancing the performance of low-light image enhancement and ensuring the quality of the enhanced images. The reflection map was input into the restoration subnet for detail restoration and noise reduction, effectively addressing the degradation and noise issues of the reflection map in low-light images. The outputs of the adjustment and restoration subnets were fused to obtain the enhanced feature map. Then, to accomplish downstream recognition tasks and 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 strategy was introduced during training to generate more challenging samples, addressing the ill-posed problem while reducing alignment errors caused by low-light images during the face detection phase and further improving network performance. Experiments on CASIA-FaceV5, YaleB, and SoF low-light face datasets demonstrate that the recognition rate of LLANet reach 94.67%, 98.22%, and 97.24%, respectively. Compared with AroFace, the recognition rate improved by 2.14, 1.58, and 2.10 percentage points on the three datasets, respectively. The results indicate that LLANet achieves high recognition accuracy in low-light scenarios.

Key words: Low-light Image Enhancement, Face Recognition, Attention Mechanism, Feature Fusion, Adversarial Data Enhancement

摘要： 现实场景中图像容易受到外部光线条件或相机参数的影响，导致图像出现整体亮度过低、视觉效果不好、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场景下的自适应人脸识别网络LLANet (Low Light Adaptive Face Recognition Network)。LLANet由4个部分组成，分别为分解子网络、恢复子网络、调节子网络和主干网络。该网络以低光图像和正常光图像为输入，首先，根据Retinex理论将它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，不仅提升低光照图像增强性能，还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，有效解决低光图像反射图退化及噪声问题，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并且防止光照特征过拟合及人脸特征提取不准确，采用加权特征融合策略，将主干网络提取原始人脸特征与增强后的特征图进行融合，获得信息更丰富的特征图。最后，引入对抗性数据增强策略，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、YaleB和SoF这3个低光照人脸数据集上进行验证，LLANet的识别率分别达到了94.67%、98.22%和97.24%。与AroFace相比，在3个数据集上的识别率分别提高了2.14、1.58和2.10个百分点。可见，LLANet在低光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