PaperlayerView exact revision on arXiv

Source-linked AI summary

11K Hands: Gender recognition and biometric identification using a large dataset of hand images

Mahmoud Afifi

arXiv:1711.04322v9cs.CV

TL;DR

The paper tackles gender recognition and biometric identification from hand images, addressing the need for a large dataset with detailed ground truth. It proposes a two-stream CNN for gender recognition and uses its features with SVM classifiers for identification. The authors report that dorsal hand images provide distinctive features for both tasks, and the dataset and implementation materials are made available for replication.

  • Problem

    Hand images contain gender-relevant and person-distinctive information, but the paper seeks a large, detailed dataset to support both gender recognition and biometric identification.

  • Method

    The paper builds a dataset of dorsal and palmar hand images, trains a two-stream CNN for gender recognition, and uses the trained CNN as a feature extractor for SVM identification classifiers.

  • Results

    Dorsal hand images outperform palmar images for gender recognition and biometric identification, while the two-stream CNN outperforms AlexNet and GoogleNet for gender classification.

  • Takeaways & Limitations

    Regular-camera dorsal hand images provide effective distinctive features for both gender recognition and biometric identification.

  • Takeaways & Limitations

    Hand images containing accessories are excluded from training to avoid visual-information bias in both tasks.

Abstract

from arXiv · show

The human hand possesses distinctive features which can reveal gender information. In addition, the hand is considered one of the primary biometric traits used to identify a person. In this work, we propose a large dataset of human hand images (dorsal and palmar sides) with detailed ground-truth information for gender recognition and biometric identification. Using this dataset, a convolutional neural network (CNN) can be trained effectively for the gender recognition task. Based on this, we design a two-stream CNN to tackle the gender recognition problem. This trained model is then used as a feature extractor to feed a set of support vector machine classifiers for the biometric identification task. We show that the dorsal side of hand images, captured by a regular digital camera, convey effective distinctive features similar to, if not better, those available in the palmar hand images. To facilitate access to the proposed dataset and replication of our experiments, the dataset, trained CNN models, and Matlab source code are available at (https://goo.gl/rQJndd).

1 Introduction

The paper addresses hand-based gender recognition and biometric identification by proposing a large hand-image dataset and a CNN-based pipeline. Its contributions include a two-stream CNN for gender classification and SVM classifiers for identification.

  • Motivation: Hand dimensions provide distinctive features that can reveal gender and serve as soft biometric information alongside primary biometric traits.Hand images also offer controlled positioning and less variability than face images.
  • Motivation: Traditional hand-biometric systems extract handcrafted features before classification or similarity measurement, whereas CNNs have increasingly outperformed such methods.The paper notes CNN susceptibility to overfitting as a drawback.
  • Contributions: The paper proposes a dataset containing more than 11,000 hand images for gender recognition and biometric identification.
  • Contributions: A two-stream CNN is presented for gender classification and is reported to outperform AlexNet and GoogleNet classification accuracies.
  • Contributions: The trained CNN supplies features to a set of SVM classifiers for biometric identification.

2 Related work

Prior hand-based gender-recognition and biometric-identification methods primarily use handcrafted geometric, local, or global features with classifiers or similarity measures. The paper focuses on regular scanned or camera-captured hand images rather than infrared hand-vein patterns.

  • Scope: Earlier methods extracted local and global handcrafted features from palmar hand images for gender recognition and biometric identification.
  • Scope: The related-work scope excludes methods based on hand-vein patterns captured with infrared imaging.
  • Prior methods: Previous studies used palm and finger geometries, Fourier descriptors, skin texture patterns, and polynomial smooth SVMs across relatively small hand-image datasets.Examples include 400 palmar images from 40 subjects and 1,920 dorsal images from 160 hands.
  • Prior methods: Hand-biometric systems commonly extract geometrical or local features with generic or customized extractors, then use similarity metrics or off-the-shelf classifiers for final decisions.Reported extractors include Radon transform, Gabor filters, SIFT, and customized shape or geometric descriptors.

3 Background

The background formulates image classification as loss minimization and motivates CNNs as nonlinear feature learners. CNN features can then be reduced and supplied to conventional classifiers such as SVMs.

  • Classification formulation: Image classification minimizes a loss function with respect to the classifier parameters so estimated labels approach true labels and misclassification decreases.
  • CNN background: CNNs learn parameters across stacked layers to reduce classification error and produce discriminative features at the final fully connected layer.
  • Feature classification: SVM classifiers can be trained on extracted CNN features without retraining the CNN and without requiring as much training data as CNN training.
  • Feature classification: CNN feature vectors can be used instead of, or alongside, handcrafted features as inputs to off-the-shelf classifiers.
  • Feature classification: Reducing CNN-feature dimensionality before classifier training can boost classification accuracy.The background gives a 1 × 4,096 AlexNet feature vector as an example of a reducible representation.

4 The proposed method

The proposed method uses guided-filter preprocessing and a two-stream CNN to learn global and local hand features for gender recognition and biometric identification.

  • 4.1 Pre-processing: Guided filtering produces a low-frequency smoothed image and a normalized high-frequency detailed layer, both resized to 224 × 224 pixels for CNN input.The detailed layer is obtained by pixel-wise division of the original image by the smoothed image, with ε used for numerical stability.
  • 4.3 LBP and CNN-features: Biometric identification combines SVM outputs trained on stream-specific, fusion-layer, and local binary pattern features using a sum rule.The trained CNN acts as a feature extractor, and one-against-all SVM classifiers are used for biometric identification.
  • 4.2 The CNN architecture: The two-stream CNN learns global features from the guided-filter output and local features from the detailed layer before combining both streams in a fusion layer.The global stream receives features such as hand shape and skin color, while the second stream processes local detail.
  • 4.2 The CNN architecture: Each stream uses AlexNet-based convolutional layers with added fully connected layers that implicitly reduce feature dimensionality through learnable parameters.The network applies activation and dropout layers to the added fully connected outputs, then appends a fusion fully connected layer.
  • 4.3 LBP and CNN-features: For gender recognition, the authors evaluate a softmax CNN and an SVM using concatenated CNN features; the latter also supports biometric identification.The concatenated vector uses features from fc9 of the first stream, fc10 of the second stream, and the fusion fully connected layer.

5 The proposed dataset

The 11K Hands dataset contains over 11,000 standardized hand images from 190 subjects, with detailed demographic, hand-side, and image-condition metadata. The experiments use subject-disjoint gender-recognition splits and repeated identification trials across different subject counts.

  • 5 The proposed dataset: The 11K Hands dataset contains 11,076 images from 190 subjects aged 18–75, covering both dorsal and palmar sides of the right and left hands.Images are 1600 × 1200 pixels and were captured against a uniform white background at approximately the same camera distance.
  • 5 The proposed dataset: Each image records subject ID, gender, age, skin color, hand side, dorsal or palmar orientation, and indicators for accessories, nail polish, or irregularities.This metadata supports detailed ground-truth information for the proposed tasks.
  • 5 The proposed dataset: Frame redundancy was reduced using structural similarity, and blurry, cropped, or oversaturated images were removed after visual checks by three people.The resulting collection averages 58 images per subject: 30 dorsal and 28 palmar.
  • 5 The proposed dataset: Images containing accessories were excluded from training to avoid bias from accessory-related visual information in gender recognition and biometric identification.The exclusion applies to the training data for both problems.
  • 5 The proposed dataset: Gender-recognition data were balanced by sampling 1,000 dorsal images per gender for training and 500 per gender for testing, with subject-disjoint train and test sets.The same sampling procedure was applied to palmar images, and experiments were repeated.
  • 5 The proposed dataset: Biometric identification uses 10 training and 4 testing images per hand side for 80, 100, and 120 subjects, repeated 10 times with random selections.Average identification accuracy, also called correct identification rate, is the evaluation metric.

6 Experimental results

Experiments evaluate the proposed two-stream CNN and downstream SVM classifiers for gender recognition and biometric identification on hand-image datasets. Results compare feature types, hand sides, classifier designs, error metrics, and processing time.

  • Experimental setup: The experiments compare handcrafted descriptors, CNN architectures, and the proposed two-stream CNN for gender recognition on the 11K Hands dataset.The tested handcrafted frameworks include BoW and Fisher vector methods, while CNN baselines include AlexNet, VGGNet, and GoogleNet.
  • Evaluation: The dataset evaluation reports average gender-recognition accuracy and biometric-identification accuracy across dorsal and palmar hand images.Table 3 covers gender recognition, while Table 4 reports identification accuracy for different numbers of subjects.
  • Gender recognition: All tested classifiers achieve higher gender-recognition rates with dorsal hand images than with palmar images.This comparison includes the proposed method and the evaluated generic classifiers.
  • Biometric identification: The ensemble of four SVM classifiers outperforms a single SVM using the combined CNN feature vector.The CNN features are used as inputs to the SVM-based biometric-identification classifiers.
  • Cross-dataset evaluation: The proposed method also achieves good IITD identification accuracy without retraining the CNN, despite differences in hand position, background, and lighting between datasets.The reported results approach those of customized handcrafted feature extractors.
  • Error analysis: Biometric error analysis uses FAR, FRR, EER, and ROC curves across thresholds, with ROC evaluation reported for 80 subjects and both hand sides.The system assigns a subject only when its SVM score exceeds threshold t; otherwise, the image is rejected.

7 Conclusion

The paper introduces the 11K Hands dataset and a two-stream CNN whose features support gender recognition and biometric identification. Across the evaluated methods, dorsal hand images generally produce better or comparable results than palmar images.

  • Method: The proposed two-stream CNN is trained for gender classification and then used as a feature extractor for SVM-based biometric identification.This links the gender-recognition model to the identification pipeline.
  • Results: Dorsal hand images yield better gender-recognition accuracies than palmar images across the evaluated frameworks and CNNs.The conclusion names BoW, Fisher vector, AlexNet, VGG, and GoogleNet among the compared methods.
  • Results: The proposed method achieves better biometric-identification accuracy with dorsal images and consistently attains better or on-par accuracy in the reported experiments.The conclusion summarizes the paper's cross-task comparison between dorsal and palmar hand images.
Loading 1711.04322v9…