Range image recognition based on statistical multiresolution approach

This paper proposes a robust multiresolution approach to detecting the structure of a noisy range image. It is assumed that the original image, consisting of planar and quadratic surfaces, is corrupted by heavy noise composed of Gaussian background noise and impulse noise. A basic principle of our approach is the exploitation of the structure of scan lines for detecting the image structure. Thus, the main part of our algorithm is essentially one-dimensional, allowing a significant decrease in computational complexity. A new curve recognition technique based on multiresolution hypothesis testing is suggested. This technique allows us to take into account domain knowledge and to improve the efficiency of the method. Finally, a procedure for detection, recognition, and reconstruction of surface patches is introduced. Key-Words: -Edge detection, multiresolution, range image, robust estimation