Lothar Hermes

Feature selection for support vector machines

In the context of support vector machines (SVM), high dimensional input vectors often reduce the computational efficiency and significantly slow down the classification process. In this paper, we propose a strategy to rank individual components according to their influence on the class assignments. This ranking is used to select an appropriate subset of the features. It replaces the original feature set without significant loss in classification accuracy. Often, the generalization ability of the classifier even increases due to the implicit regularization achieved by feature pruning.

Support vector machines for land usage classification in Landsat TM imagery

Land usage classification is an essential part of many remote sensing applications for mapping, inventory, and yield estimation. In this contribution, we evaluate the potential of the support vector machines for remote sensing applications. Moreover, we expand this discriminative technique by a novel Bayesian approach to estimate the confidence of each classification. These estimates are combined with a priori knowledge about topological relations of class labels using a contextual classification step based on the iterative conditional mode algorithm (ICM). As shown for Landsat TM imagery, this strategy is highly competitive and outperforms several commonly used classification schemes.

Parametric Distributional Clustering for Image Segmentation

Unsupervised Image Segmentation is one of the central issues in Computer Vision. From the viewpoint of exploratory data analysis, segmentation can be formulated as a clustering problem in which pixels or small image patches are grouped together based on local feature information. In this contribution, parametrical distributional clustering (PDC) is presented as a novel approach to image segmentation. In contrast to noise sensitive point measurements, local distributions of image features provide a statistically robust description of the local image properties. The segmentation technique is formulated as a generative model in the maximum likelihood framework. Moreover, there exists an insightful connection to the novel information theoretic concept of the Information Bottleneck (Tishby et al. [17]), which emphasizes the compromise between efficient coding of an image and preservation of characteristic information in the measured feature distributions.The search for good grouping solutions is posed as an optimization problem, which is solved by deterministic annealing techniques. In order to further increase the computational efficiency of the resulting segmentation algorithm, a multi-scale optimization scheme is developed. Finally, the performance of the novel model is demonstrated by segmentation of color images from the Corel data base.

A minimum entropy approach to adaptive image polygonization

This paper introduces a novel adaptive image segmentation algorithm which represents images by polygonal segments. The algorithm is based on an intuitive generative model for pixel intensities and its associated cost function which can be effectively optimized by a hierarchical triangulation algorithm. A triangular mesh is iteratively refined and reorganized to extract a compact description of the essential image structure. After analyzing fundamental convexity properties of our cost function, we adapt an information-theoretic bound to assess the statistical significance of a given triangulation step. The bound effectively defines a stopping criterion to limit the number of triangles in the mesh, thereby avoiding undesirable overfitting phenomena. It also facilitates the development of a multiscale variant of the triangulation algorithm, which substantially improves its computational demands. The algorithm has various applications in contextual classification, remote sensing, and visual object recognition. It is particularly suitable for the segmentation of noisy imagery.

On learning texture edge detectors

Texture is an inherently non-local image property. All common texture descriptors, therefore, have a significant spatial support which renders classical edge detection schemes inadequate for the detection of texture boundaries. In this paper we propose a novel scheme to learn filters for texture edge detection. Textures are defined by the statistical distribution of Gabor filter responses. Optimality criteria for detection reliability and localization accuracy are suggested in the spirit of Cannys edge detector. Texture edges are determined as zero crossings of the difference of the two a posteriori class distributions. An optimization algorithm is designed to determine the best filter kernel according to the underlying quality measure. The effectiveness of the approach is demonstrated on texture mondrians composed from the Brodatz album and a series of synthetic aperture radar (SAR) imagery. Moreover, we indicate how the proposed scheme can be combined with snake-type algorithms for prior-knowledge driven boundary refinement and interactive annotation.

Boundary-constrained agglomerative segmentation

Automated interpretation of remotely sensed data poses certain demands to image segmentation algorithms, regarding speed, memory requirements, segmentation quality, noise robustness, complexity, and reproducibility. This paper addresses these issues by formulating image segmentation as source channel coding with side information. A cost function is developed that approximates the expected code length for a hypothetical two-part coding scheme. The cost function combines region-based and edge-based considerations, and it supports the utilization of reference data to enhance segmentation results. Optimization is implemented by an agglomerative segmentation algorithm that iteratively creates a tree-like description of the image. Given a fixed tree level and the output of the edge detector, the cost function is parameter-free, so that no exhaustive parameter-tuning is necessary. Additionally, a criterion is presented to reliably select an adequate tree level with high descriptive quality. It is shown by statistical analysis that the cost function is appropriate for both multispectral and synthetic aperture radar data. Experimental results confirm the high quality of the resulting segmentations.

A New Contour-Based Approach to Object Recognition for Assembly Line Robots

A complete processing chain for visual object recognition is described in this paper. The system automatically detects individual objects on an assembly line, identifies their type, position, and orientation, and, thereby, forms the basis for automated object recognition and manipulation by single-arm robots. Two new ideas entered into the design of the recognition system. First we introduce a new fast and robust image segmentation algorithm that identifies objects in an unsupervised manner and describes them by a set of closed polygonal lines. Second we describe how to embed this object description into an object recognition process that classifies the objects by matching them to a given set of prototypes. Furthermore, the matching function allows us to determine the relative orientation and position of an object. Experimental results for a representative set of real-world tools demonstrate the quality and the practical applicability of our approach.

Combined color and texture segmentation by parametric distributional clustering

Unsupervised image segmentation can be formulated as a clustering problem in which pixels or small image patches are grouped together based on local feature information. In this contribution, parametric distributional clustering (PDC) is presented as a novel approach to image segmentation based on color and texture clues. The objective function of the PDC model is derived from the recently proposed Information Bottleneck framework (Tishby et al., 1999), but it can equivalently be formulated in terms of a maximum likelihood solution. Its optimization is performed by deterministic annealing. Segmentation results are shown for natural wildlife imagery.

Semi-supervised Image Segmentation by Parametric Distributional Clustering

The problem of semi-supervised image segmentation is frequently posed e.g. in remote sensing applications. In this setting, one aims at finding a decomposition of a given image into its constituent regions, which are typically assumed to have homogeneously distributed pixel values. In addition, it is requested that these regions can be equipped with some semantics, i.e. that they can be matched to particular land cover classes. For this purpose, class labels are provided for a small subset of the image data. The demand that the image segmentation respects those class labels implies that the segmentation algorithm should be posed as a constrained optimization problem.

Contextual classification by entropy-based polygonization

To improve the performance of pixel-wise classification results for remotely sensed imagery, several contextual classification schemes have been proposed that aim at avoiding classification noise by local averaging. These algorithms, however, bear the serious disadvantage of smoothing the segment boundaries and producing rounded segments that hardly match the true shapes. The authors present a novel contextual classification algorithm that overcomes these shortcomings. Using a hierarchical approach for generating a triangular mesh, it decomposes the image into a set of polygons that, in our application, represent individual land-cover types. Compared to classical contextual classification approaches, this method has the advantage of generating output that matches the intuitively expected type of segmentation. Besides, it achieves excellent classification results.