A Native imitative computer vision for a following generation human interface

点击翻译按钮获取中文摘要

It is important to combine the robustness with the real-time processing, when we build a computer-vision system for future-generation human interface. For this purpose, we must establish a new image understanding method that represents the image as a hierarchical structured information. In this research, we propose a new method multi-resolution analysis for an image pattern. We also apply the proposed method to some problems including Face Detection, which is important in the area of the application to the human interface. The results show the proposed method works efficiently with our system.For the multi-resolution analysis, we adopt a new method called "Image dipole analysis in Scale-Space". At each scale, the image is represented as a dipole network, which consist of local-maximal point and local-minimal point of the image. Moreover, the network at each scale is connected along the scale and forms a hyper network. On this hyper network, we also represent the target object (for example, human face) as the subgraph structure and we can detect the object this graph representation. Based on this image representation, we propose a new object detection method and apply this to human interface and realize a robust and real-time system.In the application to the face detection, we prepare a standard pattern of the dipole network of human face from a sample set with multi-resolution analysis. For an input image, we obtain the hyper network and match the subgraph which is the candidates of the face with standard pattern. Moreover, we analyze the candidates in detail and finally we acarise the result of detection.