Contributions to real time shape detection with applications to face detection.

详细信息   

• 作者：Wu ; Feng.
• 学历：Doctor
• 年：2006
• 导师：Schweitzer, Haim
• 毕业院校：The University of Texas
• 专业：Computer Science.
• ISBN：9780542932892
• CBH：3238615
• Country：USA
• 语种：English
• FileSize：3917868
• Pages：122

文摘

Shape-detection is a critical component in many computer vision systems. Large variety of shapes, lighting conditions, and the high dimensionality of image data make shape-detection a challenging problem. Our research is focused on developing techniques for improving both the accuracy and the speed of shape and object detection. We first describe contributions to the classic normalized template-matching technique. Based on the concept of integral images, we have developed an approximate normalized template-matching method, which can be much faster (100 times faster in typical applications) than the classic normalized template-matching method. In typical template-matching based shape-detection, matching needs to be performed at every possible template location in the image and at multiple image resolutions. We introduce the idea of local maxima and describe a selective template matching method. In this approach, matching values are computed only at selective locations, resulting in additional speed-up of the run time.;The detection and the recognition of shapes is typically performed on a small number of useful features. The selection of good features requires processing large amounts of data and is very time consuming. The standard approach to feature-selection is to compute many features from training data, and then select among them a subset that is useful for the task at hand. We have developed a fast selection method that works for linear features. Our approach is based on Principal Components dimensionality reduction on the training data. We show that it is not necessary to compute features for all the training data; instead, one needs only compute features for the principal components. Specifically, redundant features are identified using the classic QR Factorization algorithm, applied in the reduced space.;AdaBoosting is a recently developed powerful method for selecting a small set of features and constructing a classifier that uses these features. The standard AdaBoosting technique is symmetric in its treatment of the positive and the negative training examples. However, when AdaBoosting is applied in object-detection, the positive examples are significantly more important than the negative examples. For example, when AdaBoosting is applied to face detection, the positive examples are examples of human faces. It is less clear what are good examples of non-faces. We have developed an asymmetric formulation of AdaBoosting that allows a specified bias toward the positive examples that behaves better in such cases.