Face Hallucination

Face Hallucination Training Set

As our team started implementing the Face Hallucination (FH) algorithms explained in a paper by Li et al.[1] , the dataset used for the training phase was of question.

In the experiments conducted by Li et al.,  the experimental subset contained 1552 images which were selected from the GT database [2] and the FERET databases [3]. With a down-sampling factor of 4 for each dimension, the resolutions of the HR faces and the corresponding LR faces are 124 x 108 and 31 x 27, respectively. They have all been aligned using the method in [4], and normalized using the illumination–normalization technique employed in [5]. That is to reduce the effect of illumination variation, all the LR face images in the training dataset are normalized using an efficient illumination-normalization technique [7].

Xie and Lam [7] based their method on the assumption that the face is a combination of small flat facets. Then, they applied a Local Normalization (LN) model to eliminate the effects of additive and multiplicative noise. This approach saved them from unfeasible noise calculations.

The INface toolbox v2.0 for illumination invariant face recognition

While searching for a Matlab / Octave implementation for [7], we found the INface toolbox v2.1 [8, 9] which includes implementations of the majority of photometric normalization techniques in the literature. However, we weren’t quite sure how this related to [7]. We also found a Matlab method called LOCALNORMALIZE [10] which is a local normalization algorithm that uniformizes the local mean and variance of an image.

Evaluation Summary and Comparison of Different FH Methods

In a comprehensive survey to FH conducted by Wang et al. [6], they found that most image databases used for face image super-resolution were not sampled from surveillance camera videos, since the main application of face image super-resolution is face recognition or face retrieval from a monitor. The rest of the databases were sketches used for Face Sketch-Photo Synthesis (FSPS). Therefore, they suggested as a promising future direction and a task, an image database extracted from surveillance videos should be constructed that incorporates pose, illumination, expression, and view variant images.

GT Face Database example of uncropped version

During the implementation of the paper [1], my team made a mistake by believing that we had to crop the images in the Georgia Tech (GT) Face Database, since the images weren’t only of faces, but they had a lot of body and background details. However, further research showed that there is an already cropped version available on the official website [2]. The face images with the background removed using label files was available. As each image was manually labeled to determine the position of the face in the image. The set of label files were also available.

Cropped GT Face Database

Moreover, we had no idea what the FERET databases were. But as Proposed by Philips et al. [3], the FacE REcognition Technology (FERET) program database is a large database of facial images, divided into development and sequestered portions. The development portion is made available to researchers, and the sequestered portion is reserved for testing face recognition algorithms. The FERET evaluation procedure is an independently administered test of face-recognition algorithms. The test was designed to: (1) allow a direct comparison between different algorithms, (2) identify the most promising approaches, (3) assess the state of the art in face recognition, (4) identify future directions of research, and (5) advance the state of the art in face recognition.

FERET Face Database Example

Finally, I’d like to thank our project’s mentor for bringing up the idea behind the above discussion.

References:

[1] Yongchao Li, Cheng Cai, Guoping Qiu & Kin-Man Lam. Face hallucination based on sparse local-pixel structure. Journal, Pattern Recognition archive, Volume 47 Issue 3, March, 2014, Pages 1261-1270.
[2] Georgia Tech Face Database,〈http://www.anefian.com/research/face_reco.htm〉.
[3] P.J. Phillips, H. Wechsler, J. Huang, P.J. Rauss, The FERET database and evaluation procedure for face-recognition algorithms, Image and Vision Computing 16 (1998) 295–306.
[4] K.-W. Wong, K.-M. Lam, W.-C. Siu, An efficient algorithm for human face detection and facial feature extraction under different conditions, Pattern Recognition 34 (2001) 1993–2004.
[5] Y. Hu, K.M. Lam, G. Qiu, T. Shen, From local pixel structure to global image super-resolution: a new face hallucination framework, IEEE Transactions on Image Processing 20 (2011) 433–445.
[6] Nannan Wang, Dacheng Tao, Xinbo Gao, Xuelong Li & Jie Li. A Comprehensive Survey to Face Hallucination. International Journal of Computer Vision, Volume 106 Issue 1, January 2014.
Pages 9-30.
[7] X. Xie and K.-M. Lam, “An efficient illumination normalization method for face recognition,” Pattern
Recognition Letters, vol. 27, no. 6, pp. 609-617, 2006.
[8] Vitomir Štruc, “The INface toolbox v2.1 The Matlab Toolbox for Illumination Invariant Face Recognition: toolbox description and user manual,” Toolbox Description, User Manual, 2012.
[9]The INface toolbox v2.0 for illumination invariant face recognition by Vitomir Struc, 29 Jan 2010, http://www.mathworks.com/matlabcentral/fileexchange/26523-the-inface-toolbox-v2-0-for-illumination-invariant-face-recognition
[10] Local Normalization by Guanglei Xiong, 17 Aug 2005, http://www.mathworks.com/matlabcentral/fileexchange/8303-local-normalization?s_tid=srchtitle

Advertisements

One thought on “Face Hallucination Training Set

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s