PaperlayerView exact revision on arXiv

Source-linked AI summary

A Light CNN for Deep Face Representation with Noisy Labels

Xiang Wu, Ran He, Zhenan Sun, Tieniu Tan

arXiv:1511.02683v4cs.CV

TL;DR

Large-scale internet face datasets provide training data but often contain ambiguous or inaccurate labels, motivating compact face representations that can tolerate noise. The paper introduces Light CNNs with MFM feature selection and semantic bootstrapping, and reports state-of-the-art results across face benchmarks without fine-tuning while reducing computational and storage costs.

  • Problem

    Internet-collected large-scale face datasets often contain ambiguous or inaccurate labels, making robust representation learning with noisy data an important problem.

  • Method

    The framework uses MFM operations in lightweight CNN architectures and semantic bootstrapping to select features and relabel or remove noisy training images.

  • Results

    The single model with a 256-D representation achieves state-of-the-art results on various face benchmarks without fine-tuning.

  • Takeaways & Limitations

    The framework provides a faster, smaller face model that learns robust representations from noisy labeled data and has potential value for real-time face recognition systems.

Abstract

from arXiv · show

The volume of convolutional neural network (CNN) models proposed for face recognition has been continuously growing larger to better fit large amount of training data. When training data are obtained from 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. Lastly, 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. The code is released on https://github.com/AlfredXiangWu/LightCNN.

I. INTRODUCTION

The paper addresses noisy labels in increasingly large face-recognition datasets with a compact Light CNN framework. It combines MFM-based feature selection, lightweight architectures, and semantic bootstrapping to improve efficiency and face representation quality.

  • Max-Feature-Map Operation: MFM replaces ReLU-style thresholding with competitive feature-map selection to separate noisy and informative signals.The operation is introduced as a special case of maxout and is intended to preserve useful responses while suppressing others.
  • Light CNN architectures: The framework designs three Light CNN models following AlexNet, VGG, and ResNet ideas, targeting lower parameter counts, faster extraction, and reduced storage.
  • Noisy-label handling: Semantic bootstrapping uses predictions from a pretrained network to detect inconsistent labels and relabel or remove selected images.The method balances model predictions with original labels rather than rejecting labels indiscriminately.
  • Experimental results: The proposed 256-D single model achieves state-of-the-art results across large-scale, video-based, cross-age, heterogeneous, and cross-view face benchmarks without supervised fine-tuning.
  • Large-scale internet-collected face datasets improve CNN recognition but often contain massive noisy labels that can bias learned representations.

B. Noisy Label Problems

Massive noisy labels remain an open deep-learning problem, motivating MFM-based feature selection that separates informative from noisy signals while reducing representation size.

  • Massive noisy labels remain an open issue for deep-learning methods because accurate annotation is expensive.
  • Motivation and MFM design: MFM uses competitive activation to suppress selected neurons and separate noisy signals from informative signals.Unlike maxout, MFM is designed for feature suppression and light models rather than convex-function approximation.
  • MFM 2/1: MFM 2/1 combines two feature maps and outputs their element-wise maximum, producing a compact representation with 50% informative neurons.The output has N channels from an input with 2N channels.
  • MFM 3/2: MFM 3/2 inputs three feature maps and removes the minimal one element-wise, retaining 2/3 of the input information.Its gradient is binary, activating the retained feature map and suppressing the other.

B. The Light CNN Framework

The Light CNN framework uses MFM for local feature selection and sparse gradients, combining compact representations with progressively deeper architectures designed for efficiency.

  • MFM properties: MFM selects the optimal feature at each location and produces binary gradients that excite or suppress neurons during back propagation.
  • MFM properties: MFM layers provide compact representations and sparse gradients, concentrating training updates on responsive neurons.
  • Light CNN architectures: Light CNN-4 contains about 4,095K parameters and 1.5G FLOPS across four convolution and four max-pooling layers.
  • Light CNN architectures: Light CNN-9 combines five convolution layers, four NIN layers, MFM, and four max-pooling layers, totaling about 5,556K parameters and 1G FLOPS.It is described as deeper and faster than Light CNN-4.
  • Light CNN architectures: Light CNN-29 introduces residual blocks with two 3 × 3 convolutions and two MFM operations, using 12,637K parameters and about 3.9G FLOPS.
  • Design choices: The residual design removes batch normalization and uses a fully connected layer because aligned inputs preserve semantic and spatial information.

C. Semantic Bootstrapping for Noisy Labels

Semantic bootstrapping uses model prediction probabilities to relabel or remove inconsistent examples from a large noisy dataset, beginning with aligned face-image preparation.

  • Bootstrapping estimates a data distribution by resampling and inferring relabeled samples from originally labeled data.
  • The maximum conditional prediction probability determines the most convincing label for a training example.
  • The procedure trains Light CNN-9 on CASIA-WebFace, fine-tunes it on noisy MS-Celeb-1M, then uses predictions to relabel the noisy dataset.
  • Face-image alignment uses detected facial points to produce normalized face images for the training dataset.

IV. EXPERIMENTS

The experiments evaluate Light CNN models across face-recognition tasks, compare them with state-of-the-art methods, and analyze algorithmic behavior and semantic bootstrapping.

  • The evaluation covers various face-recognition tasks and includes comparisons with state-of-the-art methods.It also presents algorithmic analysis and detailed evaluation.

A. Training Methodology and Preprocessing

The experiments use CASIA-WebFace and MS-Celeb-1M for training, standardized grayscale aligned inputs, and multiple protocols spanning conventional, video-based, large-scale, and cross-domain recognition.

  • Training data and evaluation separation: CASIA-WebFace and MS-Celeb-1M are used for training, while testing excludes all five evaluation databases from training and fine-tuning.Features are directly extracted and compared using cosine similarity.
  • Preprocessing: Face images are aligned to 144×144 using five landmarks, randomly cropped to 128×128, converted to grayscale, and normalized by 255.
  • Evaluation databases: Evaluation covers large-scale, low-resolution, heterogeneous, video-based, and cross-domain face recognition or verification tasks.
  • LFW protocols: LFW testing includes standard verification, close-set and open-set identification, and BLUFR protocols with 10-fold experiments.Close-set performance uses Rank-1 identification accuracy; open-set performance uses Rank-1 DIR at 1% FAR.
  • Video protocols: YTF evaluation uses 10 splits of 250 positive and 250 negative pairs, averaging similarities from 100 randomly selected video samples.

C. Method Comparison

The Light CNN models are evaluated across diverse face-recognition benchmarks using cleaned large-scale training data and comparisons with established methods. Results show strong performance, including low-FAR verification, BLUFR, and large-scale recognition outcomes.

  • Training data: 79,077 identities and about 5,049,824 images remain in cleaned MS-Celeb-1M after semantic bootstrapping.Three Light CNN models with MFM 2/1 are trained on this cleaned dataset.
  • LFW comparison: Light CNN models outperform listed state-of-the-art methods on LFW verification under an unsupervised evaluation setting.The comparison includes DeepFace, DeepID2+, WebFace, VGG, CenterLoss, and SeetaFace.
  • Low-false-acceptance verification: 97.50% VR@FAR=0 is achieved by Light CNN-29, while competing methods remain below 70%.
  • Open-set identification: 93.62% open-set identification at low false acceptance is obtained by Light CNN-29, outperforming DeepFace, DeepID2+, SeetaFace, CenterLoss, and VGG.
  • BLUFR comparison: 98.88% TPR@FAR=0.1% and 92.29% DIR@FAR=1% are obtained on BLUFR verification and open-set identification, respectively.These results are reported as better than other state-of-the-art methods.

2) Video-based Face Recognition Databases

Light CNN-29 is compared with video-based and cross-domain face-recognition methods across challenging datasets. It achieves strong video-based results but performs below GoogleNet+NAN on Celebrity-1000 because it was not trained there and uses average pooling for aggregation.

  • YTF: 95.54% accuracy is achieved by a single Light CNN-29 model on YTF without fine-tuning, outperforming several state-of-the-art methods.YTF images have lower quality than LFW because of low resolution and motion blur.
  • Video-based comparison: 94.18% rank-1 accuracy is obtained by Light CNN-29 on the video-based benchmark, exceeding the compared state-of-the-art methods.
  • Celebrity-1000: Light CNN-29 outperforms MTJSR, DELM, Eigen-RER, and GoogleNet+AvePool on both close-set and open-set Celebrity-1000 protocols.
  • Celebrity-1000: Light CNN-29 performs worse than GoogleNet+NAN on Celebrity-1000 because it is not trained on that database and uses average pooling for aggregation.

3) Large-Scale Face Recognition Databases

Light CNN-29 performs strongly across large-scale, age-invariant, pose, illumination, and cross-modal face-recognition benchmarks, often without dataset-specific fine-tuning.

  • MegaFace: 73.75% rank-1 accuracy and 85.13% VR@FAR=10^-6 are reported on MegaFace, outperforming several listed competing systems.Google FaceNet and NTechLAB perform better, but use private training datasets and unknown preprocessing techniques.
  • IJB-A and IJB-B: 93.9% TAR@FAR=0.001 versus 93.0% is obtained on IJB-A compared with VGGFace2.Light CNN-29 is trained only on MS-Celeb-1M, whereas VGGFace2 uses SENet-50 and additional fine-tuning.
  • IJB-A and IJB-B: 87.7% TAR@FAR=1e-4 and 91.9% Rank-1 are reported on IJB-B versus 83.1% and 90.2% for VGGFace2.The comparison uses mixed-media verification and identification protocols.
  • Age variation: 98.55% is achieved on CACD-VS, outperforming the compared age-invariant algorithms and open-source models.The authors interpret this as potential robustness to age-variant face recognition.
  • Pose and illumination: About 95.0% accuracy at ±60° is achieved on Multi-PIE, while the framework captures features invariant to pose and illumination.Compared methods are trained on Multi-PIE, whereas Light CNN models are trained on MS-Celeb-1M.
  • Cross-modal recognition: NIR-VIS rank-1 accuracy improves from 95.82%±0.76% to 96.72%±0.23%, and VR@FAR=0.1% from 94.03%±1.06% to 94.77%±0.43%.The Light CNN models are not fine-tuned on CASIA NIR-VIS 2.0.

D. Network Analysis

The network analysis examines MFM, architecture choices, normalization, and efficiency, finding that competitive activation and lightweight design support generalization and practical deployment.

  • Activation functions: MFM is analyzed as a feature-selection operation whose competitive relationship differs from threshold-based activations.MFM 3/2 preserves more information than MFM 2/1 by inhibiting fewer neurons.
  • Network architectures: MFM-based Light CNN-4, Light CNN-9, and Light CNN-29 perform well across face-recognition datasets and fit AlexNet-, VGG-, and ResNet-style architectures.The analysis states that MFM is suitable for these different general CNN architectures.
  • Pooling and validation: Global average pooling lowers Light CNN-29 performance on both LFW and VAL1 because it does not preserve high-level spatial information.The comparison uses different Light CNN-29 configurations.
  • Pooling and validation: Batch normalization matches the non-BN model on LFW but is substantially worse on VAL1, whose images come from an independent offline source.The authors relate the dataset difference to the contrasting normalization behavior.
  • Computational efficiency: The Light CNN-29 is 10× smaller and about 5× faster than VGG, while remaining competitive with SeetaFace and CenterLoss on efficiency measures.The speed comparison is reported for a single-core i7-4790.
  • Computational efficiency: Light CNN-9 extracts features in about 40ms on MaPU and is described as deployable on embedded systems without precision degradation.The implementation uses floating-point calculation on MaPU.

E. Noisy Label Data Bootstrapping

Semantic bootstrapping filters and relabels MS-Celeb-1M samples using model predictions, with thresholds selected against validation performance to produce cleaner retraining data.

  • Evaluation setup: Bootstrapping is evaluated on LFW, VAL1, and VAL2 using ROC and precision-recall curves, including verification rate at fixed false-acceptance rates.VAL1 and VAL2 are independent of CASIA-WebFace and MS-Celeb-1M and include highly imbalanced evaluations.
  • First bootstrapping stage: The first relabeling stage accepts samples whose predicted label matches the original label and whose prediction probability exceeds p0.Thresholds p0∈[0.6, 0.7, 0.8, 0.9] are tested, with best performance at p0=0.7.
  • First bootstrapping stage: The first bootstrapped dataset, MS-1M-1R, contains 79,077 identities and 4,086,798 images.This dataset is used to retrain the Light CNN before the second resampling stage.
  • Second bootstrapping stage: The second stage accepts matching predictions or mismatching predictions whose probability exceeds p1, balancing mislabeled samples against insufficient resampling.Although LFW changes little across thresholds, p1=0.7 performs best on VAL1 and VAL2.
  • Second bootstrapping stage: After two bootstrapping rounds, MS-1M-2R contains 5,049,824 images for 79,077 identities and is used to retrain Light CNN-9.The method explicitly treats low thresholds as riskier and high thresholds as potentially ineffective because fewer images are sampled.

V. CONCLUSIONS

The paper concludes that Light CNN combines compact representation learning, efficient architectures, and semantic bootstrapping for robust face recognition on noisy data.

  • Contributions: MFM, small convolution kernels, Network in Network layers, and residual blocks reduce parameter space while supporting a compact face representation.The framework is inspired by neural inhibition and maxout activation.
  • Efficiency: Light CNN-29 extracts a representation in about 121ms on a single-core i7-4790 and contains 12,637K parameters.The paper presents the framework as faster and smaller than other published CNN methods.
  • Deployment: Semantic bootstrapping is proposed to handle noisy labels, and experiments across face-recognition tasks support potential value for real-time systems.The conclusion links the efficiency and recognition experiments to deployment potential.
Loading 1511.02683v4…