Face Detection and Fitting Algorithm robust to Lighting Changes|RISS 상세보기

다국어 초록 (Multilingual Abstract)

In this dissertation, against various light variations, we propose the new algorithm about face detection and fitting. First, our face detection algorithm is based on low-dimensional space-separated face detection. This is a simple and efficient algorithm not only to find face in photo but also to eliminate negative training images. Second, our face fitting algorithm will approximate as pseudo inverse matrix which moves to each other after we divide a training data as several Gaussian. In this dissertation, a proposed algorithm is approximated to a non-linear distribution as Gaussian distribution, and then is significant to model a linear. As same above, we apply a linear approximation to face detection and fitting. Each method is as follows.
First, let us know about an efficient Principal Component Analysis (PCA)-based low-dimensional space-separated face detection method using decision tree This dissertation proposes a PCA-based low-dimensional space-separated face detection method using simple but effective decision trees. The proposed algorithm approximates the Gaussian distributions as linear starting with nonlinearly distributed positive learning images. The contribution of this dissertation is the decision tree based on a low-dimensional space-separated face detection method. Generally, linear Gaussian distributions are not suitable for modelling nonlinearly distributed face data. Accordingly, while this dissertation repeatedly divides given nonlinear distributions into pairs of Gaussian distributions based on PCA from low dimensional space, the method continues dividing until they are modelled as Gaussian distributions. In addition, the process of repeatedly dividing a given Gaussian distribution into two Gaussian distributions continues to be stored in decision tree; it gives effective detection performance during face detection using the advantages of decision trees. The proposed PCA-based low-dimensional space-separated face detection method using decision trees was tested in three databases. First in the database composed of multi-racial faces, each face of four races showed 100% detection rate and 0 detection errors in 100 photos. In the Sung-Poggio database, there was a 93.0% detection rate and 0 detection errors. In the Schneiderman-Kanade database, there was a 92.1% detection rate and 0 detection errors ― quite surprising results.
Second, we proposed the illumination subspace Active Appearance Model(AAM) which is an idea to solve a face fitting problem. The AAM is a powerful tool for modeling a transformable image, and is widely used in computer vision and pattern recognition fields. Basic AAM shows a satisfactory fitting performance under constant lighting. However, basic AAM fails to produce good outcomes for face images in high lighting situations when the learning image is nonlinearly distributed. To solve this problem, this dissertation proposes the illumination subspace AAM. A method to update appearance parameters in the space between two partial spaces is first proposed, then we demonstrate that the appearance parameters of two models can be exchanged using the proposed method. We then propose a method to update the selected partial spaces parameter after expanding the appearance parameter algorithms to include the illumination subspace between models, and dynamically selecting the appearance model that minimizes errors for application to many illumination subspaces. The experiments show that the proposed algorithm (with the combination of these features) performs powerfully on the face images captured under various lighting conditions.

번역하기

상세검색

RISS 보유자료

상세검색

해외전자자료

Face Detection and Fitting Algorithm robust to Lighting Changes

부가정보

분석정보

이 자료와 함께 이용한 RISS 자료

나만을 위한 추천자료