Research On Face Detection And Recognition For Home Service Robot

With the progress of technology, the robots will go into each family to providenecessary services for human, of which the prospects of development is broad. In order toprovide more intimate and more targeted services, the home service robot should identifywho the person is firstly, and then provides him or her customized services according to hisor her habit. Of course, these should be achieved step by step.In this paper, the main point is face detection and recognition for the home servicerobot in indoor natural light conditions. The robot collected face images by camera in theindoors condition, but in the process of collecting the light conditions was changing inreal-time，so the face image under the condition of backlight may be collected, consequentlyit could be hard for face detection and recognition. In this paper, the face images collectedby the robot in the laboratory under different conditions of illumination was preprocessed,then detected the faces, and after that the face gray image after the preprocess, which candetected the faces, was cut into the pictures of size of128128. Then the face database canbe founded on the base of all of above. Aiming at solving the problem of face detection andrecognition, following work was done.1）The paper introduced the principle of Adaboost algorithm based on Haar-like featureand the training procedure of each classifier. The cascade classifier trained by this algorithmwas used to detect the face in the face images which were captured in the laboratory，theexperimental results showed that the method worked well under most of conditions ofillumination.2）To deal with the conditions of backlight, methods such as the classical histogramequalization, Gamma correction, single scale Retinex algorithm and multi scale Retinexalgorithm, were compared, and the results showed that the backlight images using thesemethods were still very difficult to be detected. Therefore an improved multi scale Retinexalgorithm was used, with a dynamic factor introduced, so as to eliminate the color cast,making it possible to detect some backlight faces. In order to improve the face detection rateunder backlight conditions, the dark channel prior method was introduced, following byoptimizing the transmission. Finally, combine the multi scale Retinex algorithm with thedark channel prior method to preprocess the face images, and then used the Adaboost algorithm based on Haar-like feature to detect the face images, in this way, the detection rateof backlight face can be greatly improved.3）Gabor filter is used to extract the face feature. In order to reduce the dimension andto achieve better classification, the Fisherface method was used in this paper. Firstly, theprincipal component analysis method was used to reduce the feature dimension of the facefeature extracted from the Gabor filter, then the linear discriminant analysis method wasused to project the feature into the Fisherface space, and after that it could be classified, inthis way the method solved the problem of small sample. The experimental results show that,the method has a good recognition rate for the face image collected in the laboratory.