The construction of high-quality panoramas by image mosaicing is an active area of research in the fields of photogrammetry, computer vision, image processing and computer graphics. To obtain a broader view of a scene than is available with a single view, it needs to scan multiple overlapping images and registers all these images into a large, high-resolution panoramic mosaic. Image mosaicing can create panoramas with almost arbitrarily resolution. The image mosaicing methods have distinct characteristics for diffirent fields. This dissertation investigates mosaicing algorithms for three kinds of applications, and it mainly includes multilayer microscopic image mosaicing method, document image mosaicing method for hand-held camera, and video mosaicing methods for surveillance. Firstly, it needs to scan large-scale microscopic images for the multilayer interconnection structure in the IC reverse analysis engineering. A fast autofocusing algorithm based on prediction is proposed for the microscopy images acquisition. It uses 3D grid graph to represent neighborhood relations of overlapping images, and constructs a global image alignment model for the multilayer image mosaics, thus, to get an accurate 3D representation of the multilayer structure. Based on the global model, it proposes a minimum cycle method to eliminate large alignment error caused by image mismatch. Secondly, it proposes an image mosaicing method for hand-held camera-captured document images. It does not need to restrict the camera position and calibrate the intrinsic/extrinsic camera parameters in advance, and allows greater flexibility than approaches using scanner or fixed-cameras. This paper corrects the lens distortion using two-view point correspondences. Then, it estimates the vanishing points and shear angle by mathematic morphology, and propose a hierarchical approach for the perspective rectification. it uses features correspondeces of all the overlapping image pairs to construct global alignment model. It obtains global consistent alignment parameters of all images using nonlinear optimization method, and eliminates error accumulation effectively. Finally, multiple static cameras or PTZ cameras are often used to monitor activity over a wide area in video surveillance system. This paper uses mixtures of Gaussian to model the background of each video, and then computes the homography with most proble background images. This approach can avoid alignment error caused by moving object and multi-model background. Moreover, this paper also proposes a key frame based video mosaicing approach for PTZ cameras. Key frames are selected based on amount of overlap and abundance of texture information, and all frames are matched to their latest neighbor key frames. Key frames often have very high alignment accuracy, and it can create accurate background model of scenes that moving object have transversed.
修改评论