RESUMO
The outbreak of the COVID-19 coronavirus epidemic has promoted the development of masked face recognition (MFR). Nevertheless, the performance of regular face recognition is severely compromised when the MFR accuracy is blindly pursued. More facts indicate that MFR should be regarded as a mask bias of face recognition rather than an independent task. To mitigate mask bias, we propose a novel Progressive Learning Loss (PLFace) that achieves a progressive training strategy for deep face recognition to learn balanced performance for masked/mask-free faces recognition based on margin losses. Particularly, our PLFace adaptively adjusts the relative importance of masked and mask-free samples during different training stages. In the early stage of training, PLFace mainly learns the feature representations of mask-free samples. At this time, the regular sample embeddings shrink to the prototype. In the later stage of training, PLFace converges on mask-free samples and further focuses on masked samples until the masked sample embeddings are also gathered in the center of the class. The entire training process emphasizes the paradigm that normal samples shrink first and masked samples gather afterward. Extensive experimental results on popular regular and masked face benchmarks demonstrate the superiority of our PLFace over state-of-the-art competitors.
RESUMO
The widespread dissemination of facial forgery technology has brought many ethical issues and aroused widespread concern in society. Most research today treats deepfake detection as a fine grained classification task, which however makes it difficult to enable the feature extractor to express the features related to the real and fake attributes. This paper proposes a depth map guided triplet network, which mainly consists of a depth prediction network and a triplet feature extraction network. The depth map predicted by the depth prediction network can effectively reflect the differences between real and fake faces in discontinuity, inconsistent illumination, and blurring, thus in favor of deepfake detection. Regardless of the facial appearance changes induced by deepfake, we argue that real and fake faces should correspond to their respective latent feature spaces. Particularly, the pair of real faces (original-target) remain close in the latent feature space, while the two pairs of real-fake faces (original-fake, target-fake) instead keep faraway. Following this paradigm, we suggest a triplet loss supervision network to extract the sufficiently discriminative deep features, which minimizes the distance of the original-target pair and maximize the distance of the original-fake (also target-fake) pair. The extensive results on public FaceForensics++ and Celeb-DF datasets validate the superiority of our method over competitors.
Assuntos
Aprendizado Profundo , IluminaçãoRESUMO
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 980193 face images of 10574 subjects for the training dataset and 36721 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.