Multiresolution Approach to Wavelets in Computer Vision

The goal of computer vision is to imitate the human ability to interpret the information content of images. An image is acquired by a video camera and a digitizer provides as output an array of 512 by 512 points called pixels. Each pixel gives the value of the local light intensity in the image. In computer vision we are developing numerical algorithms for understanding these images. For example, one would like to build a computer program which is able to recognize that image 8(a) is the portrait of a woman with a hat. Since the work of Rosenfeld and Thurston [1] several researchers have shown that multiresolution approaches to images provide efficient strategies for computer vision algorithms. An image can be interpreted as a sum of details which appear at different resolutions. Such a multiresolution decomposition is meaningful because to each resolution corresponds a different type of structure in the image. At a coarse resolution these details will correspond to borders of large structures like the hat of image 8(a) whereas at a finer resolution these details will rather provide texture information like in the hairs of the woman.