Kimmo Valkealahti

Co-occurrence map: quantizing multidimensional texture histograms

Co-occurrence matrices, multidimensional co-occurrence histograms, and histograms reduced by vector quantization with the self-organizing map were compared in the classification of monochrome and color textures. Increasing the histogram dimensionality improved the classification. The highest accuracy was obtained with the reduced histograms.

Experiences from operational cloud classifier based on self-organizing map

A new operational system to interpret satellite images is represented. The described method is adaptive. It is trained by examples. In the reported application a combination of textural and spectral measures is used as a feature vector. The adaptation or learning of the extracted feature vectors occurs by a self-organizing process. As a result a topological feature map is generated. The map is identified by known samples, examples of clouds. The map is used later on as a code book for cloud classification. The obtained verification results are good. The represented method is general in the sense that by reselecting features it can be applied to new problems.

Reduced multidimensional histograms in color texture description

We (1998) have developed methods for the description of monochrome and color textures with models of multidimensional co-occurrence distributions. The models are histograms of quantized multidimensional co-occurrence vectors obtained using the code words of vector quantizer as indexes of histogram bins. In the present study, the color texture analysis is further developed by selecting the co-occurring color components and the number of code vectors to minimize the classification error. The genetic algorithm is used for the optimization, and the iterative searches for the best parameters are performed by a vector quantizer with a short training time: the two-stage vector quantizer. The reduced multidimensional color histograms of 2-by-2-pixel values provide significantly higher classification accuracies than two- or three-dimensional histograms of intra- and interpixel co-occurrences. They also performed better than a Markov random field model.

Feature Selection with Self-Organizing Feature Map

Feature selection is an important phase in pattern recognition. It has also an important role in neural network methods even though it is often neglected. The fundamental function of feature selection is to find a set of features that will represent the pattern vector in a most optimal way. Only the information that is either redundant or irrelevant to the classification task is removed from the pattern vectors. The dimension of the feature vector is usually smaller than the dimension of the pattern vector so the computation time and the memory requirements are greatly reduced.

Compressing higher-order co-occurrences for texture analysis using the self-organizing map

Texture analysis is useful in many computer vision applications. One of the most useful texture feature sets is based on second-order co-occurrences of gray levels of pixel pairs. An extension of the co-occurrences to higher orders is prevented by the large size of the multidimensional arrays. We quantize the higher-order co-occurrences by the self-organizing map, called the co-occurrence map, which allows a flexible two-dimensional representation of co-occurrence histograms of any order. Experiments with natural gray level and color textures show that the method is effective in texture classification and segmentation.

Local Independent Component Analysis by the Self-Organizing Map

We introduce a neural network for the analysis of local independent components of an input signal. The network is a modification of Kohonens adaptive-subspace self-organizing map. The map units consist of weight matrices adapted to represent linear transformations which locally minimize statistical dependence among pattern vector components. Training of the map is carried out in episodes comprising pattern vectors sampled from adjacent time instants or spatial locations. The use of episodes produces independent directions which are preserved in translations of the input signal. The independent components modeled by each map unit are estimated with a nonlinear Hebbian-like learning rule, which searches for weight vectors maximizing a measure of non-Gaussianity of the scalar product of weight and pattern vectors. For demonstration, the method was applied to the segmentation of a composition image of four periodic texture fields. The spatial convolution masks, created by the map for the extraction of independent components, represent distinct frequences of particular directions.

Operational Cloud Classifier Based on the Topological Feature Map

Recently, an adaptive method to partition and interpret satellite images for cloud classification has been presented. In this application, texture measures are used as features and the segmentation and classification is based the self-organizing process. The cloud classification system has been in operational use since September 1991. In this paper, performance characteristics obtained so far are presented.

Optimal Texture Feature Selection for the Co-Occurrence Map

Textures can be described by multidimensional co-occurrence histograms of several pixel gray levels and then classified, e.g., with nearest-neighbors rules. In this work, multidimensional histograms were reduced to two dimensions using the Tree-Structured Self-Organizing Map, here called the Co-occurrence Map. The best components of the co-occurrence vectors, i.e., the spatial displacements minimizing the classification error were selected by exhaustive search. The fast search in the tree-structured maps made it possible to train about 14 000 maps during the feature selection. The highest classification accuracies were obtained using variance-equalized principal components of the co-occurrence vectors. Texture classification with our reduced multidimensional histograms was compared with classification using either channel histograms or standard co-occurrence matrices, which were also selected to minimize the classification error. In all comparisons, the multidimensional histograms performed better than the two other methods.