A multilayer self-organizing feature map for range image segmentation

-This paper proposes and describes a hierarchical self-organizing neural network for range image segmentation. The multilayer self-organizing feature map (MLSOFM), which is an extension of the traditional (singlelayer ) self-organizing feature map ( SOFM) is seen to alleviate the shortcomings of the latter in the context of range image segmentation. The problem of range image segmentation is formulated as one of vector quantization and is mapped onto the MLSOFM. The M L S O F M combines the ideas of self-organization and topographic mapping with those ofmultiscale image segmentation. Experimental results using real range images are presented. Keywords--Range image segmentation, Self-organizing feature map, Neural networks, Computer vision. 1. I N T R O D U C T I O N The availability of fast, accurate, reliable, and economical range sensors has prompted a rapid increase in the use of range images as input data for computer vision systems in recent years. A range image is usually formatted as an array of pixels such that the pixel values encode the depths or the distances of points on a visible scene surface from the range sensor. The depth value at each pixel reflects (1) the surface geometry and viewing geometry in terms of the distance of the corresponding point on a visible scene surface from the range sensor, and (2) the characteristics of the range sensor such as spatial resolution, range resolution, dynamic range, and the sensor noise parameters. The most attractive feature of using range images is that the surface information is made explicit. Although surface information can also be inferred from intensity images, it is a more difficult problem. Since a large number of factors, such as surface geometry, surface reflectance, surface texture, scene illumination, etc., are encoded in the pixel brightness value during the intensity image Acknowledgements: The authors wish to thank the anonymous referees for their insightful and detailed comments on previous versions of our paper. Their comments and suggestions have made the final paper much more readable than the earlier versions. The authors also wish to thank the Pattern Recognition and Image Processing Laboratory at Michigan State University, East Lansin~ MI for range images from their range image data base. Requests for reprints should be sent to Jean Koh, Department of Electrical and Computer Engineering, 121 Link Hall, Syracuse University, Syracuse, NY 13244-1240. formation process, techniques for deriving 3-D structure from 2-D images such as shape from shading, shape from texture, and shape from motion tend to be illposed and need to make constraining assumptions about the scene and imaging parameters. Computer vision can be looked upon as an information processing activity that involves construction of representations at successive levels of abstraction (Marr & Nishihara, 1978). A segmented image, produced by grouping the elements of an input image into semantically meaningful entities, is generally considered to be the highest domain-independent abstraction of th~ input data. Typically, a segmented image is the input to high-level vision which then utilizes domain-specific knowledge to interpret and analyze the image contents. Although depth information is explicitly available in a range image, the problems of 3-D segmentation and 3D feature extraction still need to be addressed as they do for intensity images. In the context of range images, the problem of segmentation could be looked upon as one of grouping range image pixels into clusters that represent smooth surface regions bounded by surface discontinuity contours. The purpose of this paper is to describe a neural network structure and the associated learning procedure suitable for the task of range image segmentation. The segmentation technique described in this paper is based on feature vector clustering and is mapped onto the proposed network that consists of multiple layers. Each layer is a conventional (single-layer) self-organizing feature map (SOFM) consisting of Kohonen units. The overall structure is in the form of a pyramid, thus