Font Size: a A A

Face Detection And Super-Resolution

Posted on:2009-12-21Degree:MasterType:Thesis
Country:ChinaCandidate:H X LiuFull Text:PDF
GTID:2178360242976684Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
The research of human face is one of classical problems in many fields such as pattern recognition, computer vision and computer graphics. Human face is a very complicated, multidimensional and nonrigid pattern, which has very complex physiological structure. What's more, due to people's familiarity with and sensitivity to the human face, it is comparatively difficult to do research of the human face.The first step of face processing system is to detect the position of face in the image. However, the detection of face from an image is a challenge because of the variation of face's scale, position, direction and pose, and the variation of countenance, shelter and illumination condition. In some video streams, the face is too small to track. Accordingly, we can magnify the video stream by the image super-resolution, and then track the small face. To do this task, we need to research the face image super-resolution in the static image.This thesis will explore the two key problems of human face research mentioned above. Meantime, we will propose novel algorithms to improve current methods.⑴Improve the method based on Adaboost algorithm for face detection. AdaBoost is an aggressive learning algorithm which produces a strong classifier by choosing visual features in a family of simple classifiers and combining them linearly. This kind of classifier is capable of processing images rapidly while having high detection rates. However, the algorithm has some deficiencies: the speed of training is rather slow, and the quality of the final detection depends highly on the consistence of the training set. In this thesis, a novel method based on Adaboost algorithm for face detection is presented, which includes following several ways. On one hand, the feature reducing is applied to accelerate the pace of training, on the other hand, many strategies including samples augmenting and multi-resolution searching are used to improve the performance of detecting. The problem about effect of shelter, rotation and illumination on face detection can hence be solved to some extent. Final experiment results show the proposed approach improves both the training speed and the detection efficiency. ⑵Propose the face image super-resolution algorithm which combines the global feature with local detail information. To predict the high resolution face image from the low resolution face image is a very challenging task, because the low resolution face image has lost so much detail information. We propose a sample learning based two-phase face image super-resolution method. In the first phase, we will adopt steerable pyramid to reconstruct global high resolution face image; in the second phase, we will adopt residual face image synthesis technology to compensate the local detail information of global high resolution face image. Final experiment results show our algorithm can enhance the low resolution face image into the high resolution one while the visual effect is relatively good.
Keywords/Search Tags:face detection, Adaboost algorithm, face image super-resolution, steerable pyramid, residual image
PDF Full Text Request
Related items