A Light CNN for Deep Face Representation With Noisy Labels

By Xiang Wu, Ran He, Zhenan Sun, Tien-Ping Tan. IEEE Transactions on Information Forensics and Security, 2015

1,144 citations recorded by the source corpus.

The volume of convolutional neural network (CNN) models proposed for face recognition has been continuously growing larger to better fit the large amount of training data. When training data are obtained from the Internet, the labels are likely to be ambiguous and inaccurate. This paper presents a Light CNN framework to learn a compact embedding on the large-scale face data with massive noisy labels. First, we introduce a variation of maxout activation, called max-feature-map (MFM), into each convolutional layer of CNN. Different from maxout activation that uses many feature maps to linearly approximate an arbitrary convex activation function, MFM does so via a competitive relationship. MFM can not only separate noisy and informative signals but also play the role of feature selection between two feature maps. Second, three networks are carefully designed to obtain better performance, meanwhile, reducing the number of parameters and computational costs. Finally, a semantic bootstrapping method is proposed to make the prediction of the networks more consistent with noisy labels. Experimental results show that the proposed framework can utilize large-scale noisy data to learn a Light model that is efficient in computational costs and storage spaces. The learned single network with a 256-D representation achieves state-of-the-art results on various face benchmarks without fine-tuning.

Subfields: Computer Vision and Pattern Recognition, Signal Processing

View Paper: https://www.semanticscholar.org/paper/f7a7f27516a219122000b9b915289b7127dca2f2

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Signal Processing

Citations in the corpus

ArcFace: Additive Angular Margin Loss for Deep Face Recognition (2018, background): “Although MS-Celeb-1M and MegaFace have a significant number of identities, they suffer from annotation noises [47] and long tail distributions [50].”
Character Matters: Video Story Understanding with Character-Aware Relations (2020, positive): “We utilize a state-of-the-art face detector (Zhang et al. 2017) to localize faces in each frame and extract their 256-dimensional visual features f ∈ R 256 using LightCNN (Wu et al. 2018), considering its effectiveness at general face identiﬁcation, i.e., identifying different faces of the same pers”
MobileFaceNets: Efficient CNNs for Accurate Real-time Face Verification on Mobile Devices (2018, methodology): “It is a natural choice to use the output of the global average pooling (denoted as GAPool) layer as a face feature vector, which achieves inferior verification accuracy in many researchers’ experiments [14, 5] as well as ours (see Table 2).”
Deep Transfer Learning for Person Re-Identification (2016, methodology): “With these pseudo labels, a semantic bootstrapping strategy [22] is used to train the deep model, in which the base network will”
Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis (2017, methodology): “The experiments are conducted by ﬁrstly extracting deep features with Light-CNN [38] and then compare Rank-1 recognition accuracy with a cosine-distance metric.”
Channel Augmented Joint Learning for Visible-Infrared Recognition (2021, methodology): “The lightCNN model [44] is served as a powerful baseline for VI-FR [53].”
Feature Transfer Learning for Face Recognition With Under-Represented Data (2019, methodology): “Due to label noise, we adopt a cleaned version from [43] and remove the classes overlapped with LFW and IJB-A, which results in 4 .”
Data Uncertainty Learning in Face Recognition (2020, background): “f label noise — examples have erroneously been given the labels of other classes within the dataset. Some works explore the inﬂuence of label noise [39] and how to train robust FR models in this case [17, 44, 29]. Yu et al. [50] claims in person Re-ID that another image noise brought by poor quality”
Synthetic Data for Deep Learning (2019, methodology): “Gupta et al. [224] train a CNN to detect and segment object instances for 3D model alignment with synthetic data with renderings of synthetic objects.”
AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks (2021, methodology): “Second, we propose a mechanism referred to as “ max graph operation ” (MGO) that mimics the max feature map [17].”
Probabilistic Face Embeddings (2019, methodology): “Here, we evaluate the proposed method on two different network architectures for face recognition: CASIA-Net [48] and 29-layer Light-CNN [43].”
Deep Models and Shortwave Infrared Information to Detect Face Presentation Attacks (2020, methodology): “in each channel, with a pre-trained Light CNN model [39] before being trained for 50 epochs.”
paGAN (2018, positive): “We use the pre-trained model of Light CNN[Wu et al. 2015] to compute a 256 dimension feature vector encoding the identity of the subject’s face, and employ the following identity preserving loss to enforce the identity-likeness between I0 and the generated image:”
Deep Learning for Face Recognition: Pride or Prejudiced? (2019, methodology): “Network Details: Four deep learning based face recognition networks are analyzed to answer the two key questions pertaining to the effect and existence of bias due to race and age: (i) LightCNN-9 [36], (ii) Light CNN-29 [36], (iii) ResNet50 [17], and (iv) SENet50 [19].”
Unsupervised label noise modeling and loss correction (2019, background): “The heavy lines represent the median losses and the shaded areas are the interquartile ranges. to ensure useful visual representations can be learnt (Jiang et al., 2018b; Wang et al., 2018a; Wu et al., 2018; Jiang et al., 2018a; Zlateski et al., 2018).”
WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition (2021, background): “The key engine of recent face recognition consists of network architecture evolution [31, 58, 52, 23, 28, 53, 22, 86, 24, 76, 56], a variety of loss functions [59, 47, 57, 54, 73, 14, 45, 67, 33, 32, 68, * These authors contributed equally to this work.”