Multi-scale models in image analysis (pyramids, computer vision, processing)
Ralph L. Hartley · 1984
When processing an image it is often necessary to take into account information that is present at many different scales. A structure, called a pyramid, which consists of a stack of images at different resolutions is well adapted to this situation. Each cell in the pyramid contains a model of a region of the image. Cells on the bottom level model single image pixels, and the one cell on the top level models the entire image. Each cell is also associated with an independent process, which communicates with only a small neighborhood in the pyramid. This dissertation demonstrates how, with the correct choice of the model that is located in each cell and of the process by which the models interact, accurate, robust and efficient analyses of images can be obtained in a variety of domains. If the model in each cell of the pyramid consists of an ideal step edge corrupted by noise, this approach yields a Hueckel type edge detector which efficiently detects edges at all scales. If the model consists of edges or curves which meet at corners the approach gives a hierarchical algorithm for analyzing noisy boundaries. If the model is a collection of dots within the region represented by the cell then Stevens's algorithm, which simulates human performance in analyzing Glass patterns, can be efficiently implemented in parallel. If the model is a polynomial fit to a flow field over the region represented by a cell, then a robust segmentation of optical flow fields is possible. Finally, texture discrimination can be done by using information from many different scales.