During COVID-19 coronavirus epidemic, almost everyone wears a mask to prevent the spread of virus. It raises a problem that the traditional face recognition model basically fails in the scene of face-based identity verification, such as security check, community visit check-in, etc. Therefore, it is imminent to boost the performance of masked face recognition. Most recent advanced face recognition methods are based on deep learning, which heavily depends on a large number of training samples. However, there are presently no publicly available masked face recognition datasets, especially real ones. To this end, this work proposes three types of masked face datasets, including Masked Face Detection Dataset (MFDD), Real-world Masked Face Recognition Dataset (RMFRD) and Synthetic Masked Face Recognition Dataset (SMFRD). Besides, we conduct benchmark experiments on these three datasets for reference. As far as we know, we are the first to publicly release large-scale masked face recognition datasets that can be downloaded for free at https://github.com/X-zhangyang/Real-World-Masked-Face-Dataset..

The existing occlusion face recognition algorithms almost tend to pay more attention to the visible facial components. However, these models are limited because they heavily rely on existing face segmentation approaches to locate occlusions, which is extremely sensitive to the performance of mask learning. To tackle this issue, we propose a joint segmentation and identification feature learning framework for end-to-end occlusion face recognition. More particularly, unlike employing an external face segmentation model to locate the occlusion, we design an occlusion prediction module supervised by known mask labels to be aware of the mask. It shares underlying convolutional feature maps with the identification network and can be collaboratively optimized with each other. Furthermore, we propose a novel channel refinement network to cast the predicted single-channel occlusion mask into a multi-channel mask matrix with each channel owing a distinct mask map. Occlusion-free feature maps are then generated by projecting multi-channel mask probability maps onto original feature maps. Thus, it can suppress the representation of occlusion elements in both the spatial and channel dimensions under the guidance of the mask matrix. Moreover, in order to avoid misleading aggressively predicted mask maps and meanwhile actively exploit usable occlusion-robust features, we aggregate the original and occlusion-free feature maps to distill the final candidate embeddings by our proposed feature purification module. Lastly, to alleviate the scarcity of real-world occlusion face recognition datasets, we build large-scale synthetic occlusion face datasets, totaling up to 980,193 face images of 10,574 subjects for the training dataset and 36,721 face images of 6817 subjects for the testing dataset, respectively. Extensive experimental results on the synthetic and real-world occlusion face datasets show that our approach significantly outperforms the state-of-the-art in both 1:1 face verification and 1:N face identification.

The existing masked face recognition algorithms almost tend to adopt synthetic masked face datasets for training. However, these models are limited as they rely on existing mask augmentation methods, which contain few mask patterns and cannot simulate shadows and textures in realistic scenes. To overcome this limitation, we propose a semi-supervised face recognition framework to fully exploit unlabeled real masked face samples, improving the mask robustness of the recognition model. More specifically, unlike the original face embedding network, we design a part-aware network to explore multi-region face representation based on the face structure. In this way, we obtain multiple face sub-embeddings, which correspond to different regions of the face, including the upper half, the lower half and the whole. Crucially, we use the norm of the sub-embedding to represent the activation state of the facial region features. For the input unlabeled masked face image, we restrict the sub-embedding norm of its lower half to weaken the face feature representation of the occluded area. For normal face samples, their partial features are kept activated by maintaining the sub-embedding norm, which guides the deep network does not ignore the available information. Moreover, we employ the margin-based recognition loss for normal samples to ensure that the model is sufficiently discriminative for normal facial features. Extensive experimental results on both normal and real masked face datasets show that our approach significantly outperforms the state-of-the-arts. Code is available at https://github.com/Baojin-Huang/UFace.

Due to the low computational cost of Hamming distance, hashing-based image retrieval has been universally acknowledged. Therefore, it is becoming increasingly important to quickly generate high-precision hash codes (also hash features) from images. However, the existing deep hashing methods are vulnerable to image content variations; that is, it is difficult to generate stable and consistent hash codes for similar images. In addition, generating hash codes of different lengths requires retraining the model, which is expensive in training time. To address these problems, this paper proposes a deep hashing network (DHN) with a hybrid attention mechanism and adaptive weighting (HAAW) learning. It mainly consists of a feature extraction module, feature refinement module, classification layer, hash layer and an adaptive weight layer. In particular, the hybrid attention mechanism combines bottom-up pixel saliency and top-down semantic constraints, in which the former is achieved through channel and spatial attention (CSA) and the latter is supervised by classification labels. In this way, it encourages the network to focus on dominant semantic features without being disturbed by irrelevant objects so that semantically similar images can be mapped to approximate hash codes. We further propose an adaptive weighting learning algorithm to generate weights for each bit of the hash code generated by the deep network. Then, we directly generate shorter hash codes from the available long hash code according to the importance of bits represented by the weights. This avoids retraining the network for learning hash codes of different lengths. Extensive experiments on public CIFAR-10, NUS_WIDE and ImageNet datasets show that our method has achieved substantial improvements over the counterparts in terms of precision and speed.