Adaptive multi-scale segmentation of surface data using unsupervised learning of seed positions

Abstract

This paper presents a method for multi-scale segmentation of surface data using scale-adaptive region growing. The proposed segmentation algorithm is initiated by an unsupervised learning of optimal seed positions through the surface attribute clustering with a two-criterion score function. The seeds are selected as consecutive local maxima of the clustering map, which is computed by an aggregation of the local isotropic contrast and local variance maps. The proposed method avoids typical segmentation errors caused by an inappropriate choice of seed points and thresholds used in the region-growing algorithm. The scale-adaptive threshold estimate is based on the image local statistics in the neighborhoods of seed points. The performance of this method was evaluated on LiDAR surface images.

Introduction

Surface segmentation for automatic or interactive data interpretation is encountered in many engineering applications, which deal with three-dimensional (3D) data obtained by scanning objects of interest. A typical application example, which is the subject of this paper, belongs to the LiDAR (Light Detection and Ranging) technology in remote sensing. This technology can provide a high-resolution representation of object and terrain surfaces in the form of a digital surface model (DSM), which is the height map of 3D LiDAR data (Campbell, 2002). Consequently, digital image processing techniques such as the image segmentation can be applied to segment 3D objects based on their detailed DSM images. The possibility of automatic (remote) height measurements makes this technique attractive for engineering projects connected with size measurements of particular objects of interest. Another example of surface segmentation is the automatic segmentation of range images, which are reconstructed from distance data of range scanners (Bab-Hadiashar and Gheissari, 2006). A different potential application is the industrial and medical diagnostic imaging based on ultrasound images of objects surfaces. The method presented in this paper focuses on LiDAR image segmentation of mostly natural scenes such as forests, parks and residential areas with buildings and trees.

The method of seeded region growing stands out from the multitude of image (surface) segmentation techniques due to its simplicity, time-efficient implementation and relatively good segmentation results (Adams and Bischof, 1994). Its theoretical foundation lies in the data clustering approach to image segmentation (Pal and Pal, 1993). The direct application of data clustering to image segmentation, however, does not take into account the spatial adjacency of regions and pixel connectivity, and it often results in noisy and sparse segmented regions. Advantageously, the region growing procedure as an iterative label-assignment algorithm directly analyzes region adjacency and connectivity of pixels belonging to the same region. Several algorithmic versions of the seeded region growing exist and they differ by the label-assignment procedure. For instance, the popular watershed segmentation algorithm can also be classified as a region-growing method (Beucher and Meyer, 1993). The weak side of the region growing approach is the seed-point selection, which is interactive or heuristic if based on a fully automatic procedure. Another problem, which may appear when applying a classical clustering approach to regional segmentation, is the presence of multiple regions (objects of interest) with close values of image attributes, e.g., the height attribute in the case of LiDAR data. They will be labeled as a single region, although in reality they are different isolated regions corresponding to distinct objects of interest.

Some modifications of the region-growing method were made in order to automatically select seed points. One of the existing methods consists in selecting seed locations at the regional borders using an edge detection operator (Fan et al., 2001). This procedure requires a heuristic and time-consuming post-processing to remove irrelevant seeds or merge over-segmented regions. In another similar method, a probability map is constructed to select seeds as the local minima points of an edge probability score (Mičušík and Hanbury, 2006). However, the seed selection by this method allocates multiple seeds at quasi-regular positions and requires sophisticated and time-consuming post-processing algorithms to select the optimal segmentation case. As opposed to the edge-area selection, a seed selection scheme using centers of flat areas has been proposed in Brunner and Soille (2007). A pre-segmentation by thresholding is used in the segmentation algorithm (Shan et al., 2008) to score the relevance of the obtained regions. This method is based on several existing image analysis methods, which are also present in the probability map method (Mičušík and Hanbury, 2006). A novel image segmentation scheme based on case-based reasoning was proposed to improve the performance of the watershed segmentation method (Frucci et al., 2008). It selects particular cases from a case base, which have image characteristics similar to that of the current input image. The segmentation parameters associated with the most similar case are then applied to the current image. The approach of case-base reasoning for seed selection requires a case base of segmentation outcomes, which is not always available in practice.

With regard to segmentation methods specially developed for LiDAR data, there are two major approaches. One is to process directly 3D LiDAR data and perform segmentation of surface point clouds in the 3D space (Söderman et al., 2004, Zhang et al., 2006). A significant part of the existing algorithms is based on 3D geometrical models of LiDAR surfaces, which limits their application scope to man-made or highly structured object surfaces only. Another approach, which is also adopted in this paper, is to implement a data segmentation method on the DSM image obtained from the initial LiDAR data by surface point triangulation and subsequent smoothing (Tiede et al., 2005). This method relies on the extraction of local maxima on the canopy surface to locate seed points. A different method of individual tree segmentation in forest areas is the crown delineation method based on the valley following algorithm (Leckie et al., 2003). However, this algorithm is not fully automatic since it requires a preliminary determination of forest areas and prior knowledge of key algorithm parameters. Another example is the method of gray-scale morphology used to segment treetops in a dense forest area (Andersen et al., 2001). A texture segmentation method using a seeded region growing was applied to segment coastal landforms in LiDAR imagery (Lucieer and Stein, 2005). In order to determine seed points, this method identifies image locations with minimum local variances similarly to the seed selection method (Brunner and Soille, 2007).

The main goal in the development of our segmentation method is a fully automatic selection of seed points for reliable image segmentation. Another goal is to obtain estimates for the threshold value for accurate region growing. In particular, the thresholds have to be estimated as functions of image local statistics in the neighborhoods of the seed points. The adaptive approach to surface (image) segmentation outlined in this paper can also be considered as an unsupervised learning of seed positions through local clustering of surface data with the constraint of region adjacency and pixel connectivity. The seed points are selected as the centers of local sub-clusters with high homogeneity and compactness with respect to their local inter-cluster distances.

Multi-scale image analysis is an important characteristic of the proposed approach, which provides invariance to variability of region sizes and their relative locations. The multi-scale and multi-resolution approaches to image segmentation are not new and many single-scale algorithms have been modified to implement the multi-scale image segmentation (Frucci et al., 2006, Jung, 2006, Wang, 1997, Saha et al., 2000, Duarte-Carvajalino et al., 2008). The main difference between the multi-resolution image analysis and the multi-scale approach is that the multi-scale analysis is performed once but at different local scales simultaneously, which depend on the image locations, i.e., the local scale is a function of pixel coordinates. The existing methods mostly perform the classical multi-scale image analysis based on multiple Gaussian filters at various scales. Here, a completely different scale selection principle is adopted for image segmentation, which is based on region homogeneity and its local contrast with adjacent regions. This principle substantially contributes to optimized selection of seed points and adaptive thresholds for the region growing.

The balance of this paper focuses on our multi-scale segmentation method with scale-adaptive selection of seed points and thresholds in the region growing. Section 2 describes a local clustering aspect of the conventional image segmentation, which is based on a two-criterion principle for optimized attribute clustering. The flowchart of the proposed region-growing segmentation is outlined in Section 3. Computation of local clustering map and selection of seed points are described in 4 Time-efficient computation of local clustering map, 5 Selection of seed points, respectively. Section 6 describes a modified region-growing algorithm using scale-adaptive thresholds. The experimental results of LiDAR surface segmentation are presented in Section 7 and concluding remarks are given in Section 8.