Minsoo Suk

Multiscale image segmentation using a hierarchical self-organizing map

Abstract Multiscale structures and algorithms that unify the treatment of local and global scene information are of particular importance in image segmentation. Vector quantization, owing to its versatility, has proved to be an effective means of image segmentation. Although vector quantization can be achieved using self-organizing maps with competitive learning, self-organizing maps in their original single-layer structure, are inadequate for image segmentation. A hierarchical self-organizing neural network for image segmentation is presented. The Hierarchical Self-Organizing Map (HSOM) is an extension of the conventional (single-layer) Self-Organizing Map (SOM). The problem of image segmentation is formulated as one of vector quantization and mapped onto the HSOM. By combining the concepts of self-organization and topographic mapping with those of multiscale image segmentation the HSOM alleviates the shortcomings of the conventional SOM in the context of image segmentation.

Three-Dimensional Object Recognition from Range Images

1 Introduction.- 1.1 Computer Vision.- 1.2 Three-Dimensional Object Recognition.- 1.2.1 Representation.- 1.2.2 Indexing.- 1.2.3 Constraint Propagation and Constraint Satisfaction.- 1.3 Common Goals of Three-Dimensional Object Recognition Systems.- 1.4 Qualitative Features.- 1.4.1 Study of Qualitative Properties in Low-level Vision Processes.- 1.4.2 Qualitative Features in Object Recognition.- 1.5 The Scope and Outline of the Book.- I Fundamentals of Range Image Processing and Three-Dimensional Object Recognition.- 2 Range Image Sensors and Sensing Techniques.- 2.1 Range Image Forms.- 2.2 Classification of Range Sensors.- 2.2.1 Radar Sensors.- 2.2.2 Triangulation Sensors.- 2.2.3 Sensors based on Optical Interferometry.- 2.2.4 Sensors Based on Focusing Techniques.- 2.2.5 Sensors Based on Fresnel Diffraction.- 2.2.6 Tactile Range Sensors.- 3 Range Image Segmentation.- 3.1 Mathematical Formulation of Range Image Segmentation.- 3.2 Fundamentals of Surface Differential Geometry.- 3.3 Surface Curvatures.- 3.4 Range Image Segmentation Techniques.- 3.4.1 Edge-based Segmentation Techniques.- 3.4.2 Region-based Segmentation Techniques.- 3.4.3 Hybrid Segmentation Techniques.- 3.5 Summary.- 4 Representation.- 4.1 Formal Properties of Geometric Representations.- 4.2 Wire-Frame Representation.- 4.3 Constructive Solid Geometry (CSG) Representation.- 4.4 Qualitative Representation using Geons.- 4.5 Aspect Graph Representation.- 4.6 EGI Representation.- 4.7 Representation Using Generalized Cylinders.- 4.8 Superquadric Representation.- 4.9 Octree Representation.- 4.10 Summary.- 5 Recognition and Localization Techniques.- 5.1 Recognition and Localization Techniques-An Overview.- 5.2 Interpretation Tree Search.- 5.3 Hough Clustering.- 5.4 Matching of Relational Structures.- 5.5 Geometric Hashing.- 5.6 Iterative Model Fitting.- 5.7 Indexing and Qualitative Features.- 5.8 Vision Systems as Coupled Systems.- 5.8.1 Object-Oriented Representation for Coupled Systems.- 5.8.2 Object-Oriented Representation for 3-D Object Recognition.- 5.8.3 Embedding Parallelism in an Object-Oriented Coupled System.- 5.9 Summary.- II Three-Dimensional Object Recognition Using Qualitative Features.- 6 Polyhedral Object Recognition.- 6.1 Preprocessing and Segmentation.- 6.1.1 Plane Fitting to Pixel Data.- 6.1.2 Clustering in Parameter Space.- 6.1.3 Post Processing of Clustering Results.- 6.1.4 Contour Extraction and Classification.- 6.1.5 Computation of Edge Parameters.- 6.2 Feature Extraction.- 6.3 Interpretation Tree Search.- 6.3.1 Pose Determination.- 6.3.2 Scene Interpretation Hypothesis Verification.- 6.4 Generalized Hough Transform.- 6.4.1 Feature Matching.- 6.4.2 Computation of the Transform.- 6.4.3 Pose Clustering.- 6.4.4 Verification of the Pose Hypothesis.- 6.5 Experimental Results.- 6.6 Summary.- 7 Recognition of Curved Objects.- 7.1 Representation of Curved Surfaces.- 7.1.1 Extraction of Surface Curvature Features from Range Images.- 7.2 Recognition Using a Point-Wise Curvature Description.- 7.2.1 Object Recognition Using Point-Wise Surface Matching.- 7.3 Recognition Using Qualitative Features.- 7.3.1 Cylindrical and Conical Surfaces.- 7.3.2 The Recognition Process Using Qualitative Features.- 7.3.3 Localization of a Cylindrical Surface.- 7.3.4 Localization of a Conical Surface.- 7.3.5 Localization of a Spherical Surface.- 7.3.6 An Experimental Comparison.- 7.4 Recognition of Complex Curved Objects.- 7.5 Dihedral Feature Junctions.- 7.5.1 Types of Dihedral Feature Junctions.- 7.5.2 Matching of Dihedral Feature Junctions.- 7.5.3 Pose Determination.- 7.5.4 Pose Clustering.- 7.6 Experimental Results.- 7.7 Summary.- III Sensitivity Analysis and Parallel Implementation.- 8 Sensitivity Analysis.- 8.1 Junction Matching and Pose Determination.- 8.2 Sensitivity Analysis.- 8.3 Qualitative Features.- 8.4 The Generalized Hough Transform.- 8.4.1 The Generalized Hough Transform in the Absence of Occlusion and Sensor Error.- 8.4.2 The Generalized Hough Transform in Presence of Occlusion and Sensor Error.- 8.4.3 Probability of Spurious Peaks in the Generalized Hough Transform.- 8.5 The Use of Qualitative Features in the Generalized Hough Transform.- 8.5.1 Reduction in the Search Space of Scene Interpretations due to Qualitative Features.- 8.5.2 Reducing the Effect of Smearing in Parameter Space using Qualitative Features.- 8.5.3 The Probability of Random Peaks in the Weighted Generalized Hough Transform.- 8.5.4 Determination of ?k(x), pk(x) and P(k).- 8.6 Weighted Generalized Hough Transform.- 9 Parallel Implementations of Recognition Techniques.- 9.1 Parallel Processing in Computer Vision.- 9.1.1 Parallel Architectures.- 9.1.2 Parallel Algorithms.- 9.2 The Connection Machine.- 9.2.1 System Organization.- 9.2.2 Performance Specifications.- 9.3 Object Recognition on the Connection Machine.- 9.3.1 Feature Extraction.- 9.3.2 Localization of Curved Surfaces.- 9.3.3 Computation of Dihedral Feature Junctions.- 9.3.4 Matching and Pose Computation.- 9.3.5 Pose Clustering.- 9.4 Object Recognition on the Hypercube.- 9.4.1 Scene Description.- 9.4.2 Model Data.- 9.4.3 Scene Feature Data.- 9.4.4 Pruning Constraints.- 9.4.5 Localization.- 9.5 Mapping the Interpretation Tree on the Hypercube.- 9.5.1 Breadth-First Mapping of the Interpretation Tree.- 9.5.2 Depth-First Mapping of the Interpretation Tree.- 9.5.3 Depth-First Mapping of the Interpretation Tree with Load Sharing.- 9.5.4 Experimental Results.

On machine recognition of hand-painted Chinese characters by feature relaxation

Abstract A new relaxation matching method based on features is introduced for the recognition of hand-printed Chinese characters. The types of features are selected carefully to reflect the structural information of characters. Matching probabilities between two features, one from the mask and the other from input, are computed by the relaxation method. A new distance measure between two characters based on these matching probabilities is defined. We demonstrate, through examples, the utility of the new approach in the recognition of hand-printed Chinese characters. It is especially powerful in distinguishing similarly-shaped characters within a cluster produced by preclassification.

A multi-layer Kohonen's self-organizing feature map for range image segmentation

A self-organizing neural network for range image segmentation is proposed and described. The multi-layer Kohonens self-organizing feature map (MLKSFM), which is an extension of the traditional single-layer Kohonens self-organizing feature map (KSFM), 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 MLKSFM. The MLKSFM is currently implemented on the Connection Machine CM-2, which is a fine-grained single instruction multiple data (SIMD) computer. Experimental results using both synthetic and real range images are presented.<<ETX>>

A hierarchical neural network and its application to image segmentation

The problem of image segmentation can be formulated as one of vector quantization. Although self-organizing networks with competitive learning are useful for vector quantization, they, in their original single-layer structure, are inadequate for image segmentation. This paper proposes and describes a hierarchical self-organizing neural network for image segmentation. The hierarchical self-organizing feature map (HSOFM) which is an extension of the traditional (single-layer) self-organizing feature map (SOFM) is seen to alleviate the shortcomings of the latter in the context of image segmentation. The problem of image segmentation is formulated as one of vector quantization and mapped onto the HSOFM. The HSOFM combines the ideas of self-organization and topographic mapping with those of multi-scale image segmentation. Experimental results using intensity and range images bring out the advantages of the HSOFM over the conventional SOFM.

Recognition and localization of objects with curved surfaces

This paper concerns the problem of recognition and localization of three-dimensional objects from range data. Most of the previous approaches suffered from one or both of the following shortcomings: (1) They dealt with single object scenes and/or (2) they dealt with polyhedral objects or objects that were approximated as polyhedra. The work in this paper addresses both of these shortcomings. The input scenes are allowed to contain multiple objects with partial occlusion. The objects are not restricted to polyhedra but are allowed to have a piecewise combination of curved surfaces, namely, spherical, cylindrical, and conical surfaces. This restriction on the types of curved surfaces is not unreasonable since most objects encountered in an industrial environment can be thus modeled. This paper shows how the qualitative classification of the surfaces based on the signs of the mean and Gaussian curvature can be used to come up withdihedral feature junctions as features to be used for recognition and localization. Dihedral feature junctions are robust to occlusion, offer a viewpoint independent modeling technique for the object models, do not require elaborate segmentation, and the feature extraction process is amenable to parallelism. Hough clustering on account of its ease of parallelization is chosen as the constraint propagation/ satisfaction mechanisms. Experimental results are presented using the Connection Machine. The fine-grained architecture of the Connection Machine is shown to be well suited for the recognition/localization technique presented in this paper.

Qualitative features and the generalized hough transform

Abstract In this paper we show how the use of qualitative features can enhance the performance of recognition and localization techniques, in particular, the Generalized Hough Transform. Qualitative features (i.e. scene features with qualitative attributes assigned to them) are shown to be effective in pruning the search space of possible scene interpretations and also reducing the number of spurious interpretations explored by the recognition and localization technique. The redundancy of the computed transform and the probability of spurious peaks of significant magnitude due to random accumulation of evidence are two criteria by which the performance of the Generalized Hough Transform is judged. The straightforward Generalized Hough Transform shows a high probability of spurious peaks of significant magnitude even for small values of redundancy and small magnitude of the search space of scene interpretations. The use of qualitative features enables us to come up with a weighted Generalized Hough Transform where each match of a scene feature with a model feature is assigned a weight based on the qualitative attributes assigned to the scene feature. These weights could be looked upon as membership function values for the fuzzy sets defined by these qualitative attributes. Analytic expressions for the probability of accumulation of random events within a bucket are derived for the weighted Generalized Hough Transform and compared with the corresponding expression for the straightforward Generalized Hough Transform. The weighted Generalized Hough Transform is shown to perform better than the straightforward Generalized Hough Transform. An experiment for the recognition of polyhedral objects from range images is described using dihedral junctions as features for matching and pose computation. The experimental results bring out the advantages of the weighted Generalized Hough Transform over the straightforward Generalized Hough Transform.

Sensitivity analysis for matching and pose computation using dihedral junctions

Abstract Recognition-via-localization is a popular approach in 3-D object recognition. This approach relies on the propagation of constraints that arise from the matches of local geometric features and could therefore be treated as a constraint satisfaction problem. Hough clustering, which verifies the consistency of local geometric constraints by determining the pose of the object in parameter space, is a popular technique owing to its conceptual simplicity and potential ease of parallelization. Our previous work has shown the usefulness of dihedral junctions for the recognition and localization of polyhedral objects and dihedral feature junctions for the recognition and localization of curved objects made up of piecewise combinations of conical, cylindrical and spherical surfaces. Experimental results from our previous work showed that the computed pose parameters are sensitive to the difference in the included angle between the scene and model dihedral junctions or the scene and model dihedral feature junctions. A formal analysis of the sensitivity of the computed pose to the difference in the included angle between the scene and model dihedral junctions or the scene and model dihedral feature junctions is presented in this paper. The results of the formal sensitivity analysis were found to be in conformity with the experimental results from our previous work and so the work presented in this paper could be treated as a sequel to our previous work. Based on the results of the sensitivity analysis, the rotation parameters were found to be more sensitive than the translation parameters which, in comparison, were far more robust. It is also shown how the introduction of redundancy in parameter space results in greater robustness in the computed pose. Although the analysis in this paper is based on the matching of dihedral junctions or dihedral feature junctions, the approach taken in the sensitivity analysis is general and can be applied to the matching based on other feature types.

An object-detection algorithm based on the region-adjacency graph

An object-detection algorithm based on the region adjacency graph of the segmented image is proposed. The new algorithm can detect not only simple, homogeneous objects but also more complex objects consisting of several contiguous subregions. Two examples are presented to demonstrate the usefulness of the algorithm.

Segmentation using a competitive learning neural network for image coding

This paper describes a practical segmentation procedure using a simple competitive learning neural network to yield a complete segmentation suitable for segmentation-based image coding. Image segmentation is considered as a vector quantization problem. The procedure using the FSCL neural network for the vector quantization has the two main parts: primary and secondary segmentation. In the primary segmentation, an input image is finely segmented by the FSCL. In the secondary segmentation, a lot of small regions and similar regions with larger size generated in the preceding step are eliminated or merged together by the FSCL which performs partitioning and learning every input vector. Experimental results show that the procedure described here yields the reconstructed image of reasonably acceptable quality even at the low bit rate of 0.25 bit/pel.