Meta-learning for image processing based on case-based reasoning

Abstract

We propose a framework for model building in image processing by meta-learning based on case-based reasoning. The model-building process is seen as a classification process during which the signal characteristics are mapped to the right image-processing parameters to ensure the best image-processing output. The mapping function is realized by case-based reasoning. Case-based reasoning is especially suitable for this kind of process, since it incrementally allows one to learn the model based on the incoming data stream. To find the right signal/image description of the signal/image characteristics that are in relationship to the signal-processing parameters is one important aspect of this work. In connection with this work intensive studies of the theoretical, structural, and syntactical behavior of the chosen image-processing algorithm have to be done. Based on this analysis we can propose several signal/image descriptions. The selected image description should summarize the cases into groups of similar case and map these to the same processing parameters. Having found groups of similar cases, these should be summarized by prototypes that allow fast retrieval of several groups of cases. This generalization process should permit building up the model over the course of time based on the incrementally obtained data stream. We studied this task for image segmentation based on the Watershed-Transformation. First, we studied the theoretical and the implementation aspects of the Watershed Transformation and drew conclusions for suitable image descriptions. Four different descriptions were chosen - statistical and texture features, marginal distributions of columns, rows, and diagonal similarity between the regional minima of two images, and the shape descriptor based on central moments. Our study showed that the weighted statistical and texture features and the shape descriptor based on central moments have yielded the best image description so far for the Watershed Transformation. It can best separate the cases into groups having the same segmentation parameters and it sorts out rotated and rescaled images. Generalization over cases can also be performed over the groups of case. It helps to speed up the retrieval process and to learn incrementally the general model. © 2010 Springer-Verlag Berlin Heidelberg.