Research On Eyes Location Method Based On Support Vector Machines

Computer vision is one of the most challenging fields. Since it has huge development potential, computer vision has been attracting many researchers to explore and study it in depth. With the development of computer performance and popularity of electronic product, the human-eyes location has become an important research issue in computer vision, which is very valuable in practical application such as virtual face animation, digital video processing, visual fatigue detection, human-machine interaction, visual game and so on.This thesis studies the method of "human face-human eye", a double progressive testing structure on the basis of the intensive research into related problems of eye location. It is a new method of eyes location from coarse strategy to fine strategy, which means locating eyes in the measured facial region. Thus the detectable rate will be improved greatly. The thesis focuses on the following aspects:Face image pre-processing. This thesis introduces several commonly-employed and effective methods of the eye location and then proposes a new method for eyes location in images. To reduce the false detecting rate and save storage space, some basic pre-processing methods such as image enhancement and image grayscale are used in the primarily-collected face image, which lays a solid foundation for the next work of detection and location.Detect the facial regions in images. The facial color, important information of face, does not change with different details of facial features and facial expressions. Although the facial color is different from person to person, it is stable compared with other information. Meanwhile it is obviously different from non-facial areas. Altering situations such as rotation, expression and gesture can be applied in the facial color. The skin color can be used to break up the facial region from background. So this thesis employs a facial region detecting method based on skin color. By comparing the different color space, better-color space is used to create a good skin color model, and then use the light compensation, similarity segmentation, and morphological processing steps to get the facial region. Color-based face region detection method used in the thesis has better results. The results of the experiments indicate that this method is effective.Eye location is based on support vector machine. Locating eyes in the detected facial region will narrow the scope of the human eye and reduce the dimensions of training samples. First of all, pre-process the eye and non-eye samples from different people in the complex environment. The PCA is used to reduce dimensions so that computation can be decreased greatly. Then select the appropriate kernel function and optimize the parameters. Sequential minimal optimization algorithm is introduced for off-time training and then a set of support vector and corresponding weight can be obtained. Then detect eyes in the range of the upper face using the result of SVM to complete the prelimiary detection of eyes. Finally complete the eye detection using the geometry characteristics method.The theoretical study and a large number of simulations prove that the new method of eyes location from coarse strategy to fine strategy shows a higher accuracy and robustness. It proves the effectiveness and superiority of this method.