Research On Eye-detection By Improved Adaboost Approach

Eye detection is a challenging problem in computer vision field. Due to the complexity of the real word, it is not easy to detect human eye in cluttered image. Moreover, by the impact of illumination, pose, rotation, occlusion, and so many other factors, eye detection becomes more difficult. Since AdaBoost algorithm was used by Paul Viola in face detection and achieved milestone significance, the algorithm has also been introduced into the eye detection research. Proceed from the AdaBoost algorithm, we use scattered rectangle features which are more expressive than haar rectangle features in recognition. Furthermore, a new AdaBoost cascade method is proposed by adjusting the minimum detection rate and the maximum false detection rate of each stage classifier. The major research work includes the following aspects:1. Instead of using haar rectangle feature, scattered rectangle feature has been used in our experiment to train a classifier. In order to solve the problem of explosion number of features, an optimized method has been proposed. Original haar rectangle feature can only express information in horizontal, vertical or45°direction. But, scattered rectangle feature which break the constraints of adjacent and alignment to each other can express information in almost every direction. Experimental results show that scattered rectangle feature is more expressive than original haar rectangle feature and the training time has been reduced to about one third of the original method.2. A new method that adjusts the minimum detection rate and the maximum false detection rate of each stage classifier has been proposed. In traditional cascade, the minimum detection rate and the maximum false detection rate of each stage classifier is fixed. However, in the later period of the training, it becomes very difficult to train. In order to achieve the predetermined detection rate, the stage classifier must contain more weak classifiers which are time consuming. Experiment results show that scattered rectangle feature combine with improved cascade is more effective than original haar rectangle feature combine with traditional cascade and can effectively reduce the false detection rate.3. An eye detection system has been designed using OpenCV library. OpenCV library is an open source computer vision library which encapsulates many functions that can be used in computer vision research, and it is convenient and effective. The system achieve high detection rate and also robust to complex scene like illumination, posture, rotation, occlusion, eye closed and so on. In addition, we construct a huge eye database. We cut eye regions manually in face images selected from BioID and MIT Face Database and save them in jpg file. The eye database is useful for eye detection research.