Papers Gallery

In this paper, we have proposed a new robust method of eye detection based on Gabor filter, neural network, skin color, and eye template. First, the faces are detected by a system composed of a Gabor filter, and a neural network. Then, in the faces detected, we have removed the skin regions, and keep only the non-skin regions which have a high probability to be eyes. Then, these non-skin regions are compared with an eye template to locate the position of the eyes. The obtained results show that our method is robust and provides superior performance compared to other methods recently published.
Eye detection is a very interesting area of research which allows to verify the presence of the eyes in an image and to locate their positions. In recent years, it received special attention in the computer vision community. This is mainly due to the emergence of many applications such as face recognition 1,2, video surveillance 3,4, facial expression recognition 5,6, age estimation 7, gaze tracking systems 8,9, and driver fatigue monitoring systems 10-13.
Nevertheless, several difficulties can limit the eye detection, like: the orientation and shape of eyes, the lighting conditions, and the presence or absence of structural components such as glasses. To overcome these difficulties, various techniques have been developed in recent years and can be classified into four categories 14,15: Template-matching methods 16,17, feature-based methods 18-21, appearance-based methods 22-26, and hybrid methods 27-34.
The majority of eye detection methods are based on the concept of image scanning. It is to scan the image with windows in different sizes and to classify them into eye or not-eye. This can take a long time because the detection requires a very large amount of test, and it can produce a very high number of false detection. To avoid scanning step, firstly, we proposed to determine areas likely to be the eyes, then we determine which of these are eyes.
Therefore, this article presents an eye detection method based on neural networks, skin color and eye template. It includes three algorithms. In the first, we used a simple and effective method for the detection of faces which is based on a multilayer perceptron neural network (MLP) with the retro propagation function trained with Gabor filter feature. The second algorithm, consists of removing the regions of the skin in the detected faces and the remaining non-skin regions are considered as regions likely to be eyes. In the third algorithm, the non-skin regions are compared with an eye template to locate the eyes.
This paper is organized as follows: the second part presents some related works. Our approach as well as the results obtained and their interpretations are the subject of the third and the fourth part. The conclusion and perspectives of this work are presented in part five.