Range image segmentation by dynamic neural network architecture

In this paper, a dynamic neural network architecture for jump and crease edge detection in range images is proposed. Here the weights are based on fractional differentiation whose derivative index is varied from zero to one in discrete steps. A set of edge pixels is selected as ''seed'' pixels at the start after convolving the neural edge operator at a very small derivative index. These are then linked progressively to ''real'' edge pixels extracted in subsequent stages of fractional differentiation. The robustness of this technique in identifying the crease edge pixels is demonstrated on a synthetic range image data with added noise in a precisely controlled environment. Then the technique is tested on a set of real world range image data.