Font Size

HOME > No.6, Sep 2016 > Research Highlights : Can robots recognize faces even under backlighting?

Can robots recognize faces even under backlighting?

An adaptive contrast adjustment for illumination-invariant face appearance By Jun Miura
Jun Miura (right of picture) with PhD candidate Bima Sena Bayu Dewantara
Jun Miura (right of picture) with PhD candidate Bima Sena Bayu Dewantara

Jun Miura and his colleagues have developed a novel technique to address the problem of vision-based face detection and recognition under normal and severe illumination conditions. This technique contributes to help robotic systems that use face information for providing user-dependent services to work well under a large variety of illumination conditions.

Vision-based face detection and recognition is one of the most rapidly growing research areas in computer vision and robotics and is widely used in several human related applications. However, vision-based face detection and recognition has been shown to be effective only under normal illumination conditions. In developing an algorithm for face detection and recognition, it is crucial to consider both normal and severe illumination conditions. One approach is to convert face images under various illumination conditions into ones with invariant face appearance while preserving the face-specific characteristics such as texture and facial features.

Result of the illumination invariant face processing using Yale B Extended database: input images (top) and processed images (bottom).

Now, researchers at the Department of Computer Science and Engineering at Toyohashi University of Technology have developed a novel technique to adaptively adjust the effect of lighting on human faces by employing an extended reflectance model. The model has one variable (illumination ratio), which is controlled by Fuzzy Inference System (FIS). To cope with a vast variety of illumination conditions, the FIS rule was optimized using Genetic Algorithm (GA).

Figure 2.
R Results of illumination invariant face recognition for real implementations: (a) person #1 outdoor, (b) person #1 indoor, (c) person #2 outdoor and (d) person #2 indoor. Small image in the bottom-right side of each image is the input image.

The first author PhD candidate, Bima Sena Bayu Dewantara, explained, “To eliminate the effects of light, image contrast should be adjusted adaptively. To produce an invariant face appearance under backlighting, for example, cheeks need to be brightened, while the eyeballs must be kept dark. Such an adaptive contrast adjustment can be performed using the developed reflectance model, and we have shown that a combination of Fuzzy Inference System (FIS) and Genetic Algorithm (GA) is very effective for implementing the model."

Professor Jun Miura said, "By just adding this contrast adjustment to present face recognition systems, we can significantly improve the accuracy and performance of face detection and recognition. Moreover, this adjustment runs in real-time, and therefore, it is appropriate for real-time applications such as robot and human-interaction systems.”

A face not only provides a person's identity but also provides other information such as a person’s focus of attention and the degree of tiredness. Obtaining such information is useful for smooth human–machine interaction, and researchers expect that the proposed contrast adjustment method will also be useful in various situations, especially under severe illumination conditions.

This study was partly supported by a Grant-in-Aid for Scientific Research No. 25280093 by JSPS, Japan.


Bima Sena Bayu Dewantara and Jun Miura (2016). OptiFuzz: A robust illumination invariant face recognition and its implementation, Machine Vision and Applications, DOI: 10.1007/s00138-016-0790-6.



情報・知能工学系 三浦純教授らの研究グループがさまざまな照明条件下で顔の発見と認識が可能になる新たな手法を開発しました。この手法により、顔の特徴を使って利用者を認識し、利用者に応じたサービスを提供するロボットが、厳しい照明条件下でも動作できるようになります。


そこで、本学の情報・知能工学系の研究グループ(博士後期課程学生 Bima Sena Bayu Dewanntaraおよび三浦教授)は、拡張された光の反射モデルを用いて照明の影響を適応的に調整する新たな手法を開発しました。このモデルは調整可能な一つの変数(illumination ratioと呼ぶ)を持ち、その変数をファジィ推論システムによって制御します。そして、さまざまな照明条件に対応するため、推論システムが利用するファジィ推論規則を遺伝的アルゴリズム(GA)によって最適化しておきます。



本研究成果は、平成28年7月15日(金)にMachine Vision and Applications誌上に掲載されました。


Share this story

Researcher Profile

Jun Miura
Name Jun Miura
Affiliation Department of Computer Science and Engineering
Title Professor
Fields of Research Intelligent Robotics / Robot Vision / Artificial Intelligence