Short Communication
A gender classification system robust to occlusion using Gabor features based (2D)2PCA

https://doi.org/10.1016/j.jvcir.2014.03.009Get rights and content

Highlights

  • Develop a gender classification system for both occlusion free and occluded faces.

  • In occlusion free case, it enhances the classification accuracy and speed.

  • System renders accuracies above 90% under lower occlusion conditions.

  • System also survive in higher occlusion conditions by giving minimum of 86% accuracy.

  • We analyze the impact of various face parts in the context of gender classification.

Abstract

Recognizing gender of a person from occluded face image is a recent challenge in gender classification research. This work investigates the issue and proposes a gender classification system that works for non-occluded face images to face images occluded up to 60%. Local information of the face, which carries the most discriminative features to find the gender, is gathered by dividing the face image into M×N sub-images. Subsequently, features are calculated for every sub-image by applying (2D)2PCA on each illumination invariant real Gabor space generated using Gabor filter. Support Vector Machine is used for classification. Experiments are performed on five databases. In case of non-occluded face images, the proposed approach gives 98.4% classification rate on FERET database. For occluded face images, occlusions ranging from 10% to 60%, results are quite competitive with accuracies around 90%. Present work also analyzes the impact of various face components in the context of gender classification.

Introduction

Face is a characteristic of human beings which uniquely reveals their identity, age and emotions. Gender classification from a person’s face can play an important role in a variety of computer vision based applications. Surveillance and security system design to find missing persons or criminals, biometric authentication systems such as smart systems controlling the access of people to prohibited areas, and search engines with an image filter to optimize the search are some of the potential applications. Gender classification, if used in conjunction with face recognition, makes the face recognition task twice as fast by eliminating the search for a particular gender.

Face biometrics is different from all other types of biometrics in the way face images are captured. In surveillance based security systems, a person does not volunteer himself/herself for scanning. Cameras placed in public areas usually examine thousands of faces for identification and therefore captured face images may not be perfect frontal images for examination. The probability of capturing occluded face images is very high either due to unintentional reasons (e.g. wearing sunglasses or scarf), or may be overlapping with the objects present in the environment. Also, a person may hide his face intentionally if he is aware that he is under surveillance.

In real life situations, it might be desirable to recognize gender from a face which may be partially occluded due to injury or some other reasons. In such situations a system is sought after, that is accustomed to gender recognition from occluded faces with a good classification rate. The objectives of this work are:

  • 1.

    To develop an efficient gender classification system not only for non-occluded face images but also for occluded face images.

  • 2.

    To enhance the classification accuracy and speed through an efficient and small sized feature vector.

  • 3.

    To develop a gender classification system robust against varying extents of occlusion for occluded face images.

For non-occluded face images, the proposed work gives acceptable classification rate of more than 95% on most of the face databases except on a large database collected in unconstrained environment. The performance of the system is discussed for face images occluded up to 60%.

Contents are organized as follows; Section 2 explores work done so far in gender classification; Section 3 describes the proposed gender classification system; experiments and performance analysis are covered in Section 4; and finally, Section 5 concludes the work along with the future scope.

Section snippets

Related work

Most of the researchers working on gender classification have focused on the application of a variety of feature extraction and classification techniques on full frontal face images. A number of works exist in literature targeting gender classification on occlusion free face images. Golomb et al. [1] and Tamura et al. [2] used neural network to classify gender from low resolution images. Later on, a Support Vector Machine (SVM) based method [3] was used to identify gender from low resolution

Proposed gender classification system

The proposed system starts with extraction of a Region of Interest (ROI) containing the most relevant information of face from the face image. To address the occlusion problem, local features from the face are extracted in addition to the global information of the face. It is observed that local features in face images are most robust to distortions caused by expression and illumination changes. To target these local properties, the Region of Interest (ROI) of a face image is partitioned into Si

Experiments and performance analysis

The proposed technique is exhaustively tested on five face databases for non-occluded as well as occluded face images. Analysis of experimental results gives some important clues about the relative importance of various face parts (eyes, nose, lips, etc.) for successful gender recognition.

Conclusion

A gender classification approach exploiting the potential of (2D)2PCA and Gabor filter is presented in this work. Subdivision of a face image into varying sized blocks improves the performance of the system as compared to the approaches where gender classification directly targets full face image. The assertion is empirically verified by evaluating technique on standard face databases. Performance of the proposed system also owes to preliminary AFSI extraction which in a true sense assimilates

References (37)

  • B.A. Golomb, D.T. Lawrence, T.J. Sejnowski, Sexnet: a neural network identifies sex from human faces, in: NIPS, 1990,...
  • B. Moghaddam et al.

    Learning gender with support faces

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2002)
  • W. Chuan xu, L. Yun, L. Zuo yong, Algorithm research of face image gender classification based on 2d Gabor wavelet...
  • O. Ozbudak, M. Tukel, S. Seker, Fast gender classification, in: IEEE International Conference on Computational...
  • J. Bekios Calfa et al.

    Revisiting linear discriminant techniques in gender recognition

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2011)
  • P. Rai et al.

    Gender classification techniques: a review

  • N. Sun et al.

    Gender classification based on boosting local binary pattern

  • H. Lu et al.

    Automatic gender recognition based on pixel-pattern-based texture feature

    J. Real-Time Image Process.

    (2008)
  • Cited by (0)

    View full text