Image Analysis and Automatic Surface Identification by a Bi-Level Multi-Classifier Jos´ e Mar´ ıa Mart´ ınez-Otzeta, Basilio Sierra and Elena Lazkano Dept. of Computer Science and Artificial Intelligence University of the Basque Country P. Manuel Lardizabal, s/n. 20009. San Sebasti´ an, Spain e-mail: ccbmaotj@si.ehu.es (Jos´ e Mar´ ıa Mart´ ınez-Otzeta) Web site: http://www.sc.ehu.es/ccwrobot Abstract. Combining the predictions of a set of classifiers has shown to be an effective way of creating composite classifiers that are more accu- rate than any of the component classifiers; we have performed a research work consisting of the design, development and experimental use of a multi-classifier system for image analysis and surface classification of the different segments that might appear on a given picture in order to help a Mobile Robot in its navigation task. The presented approach combines a number of component classifiers which are standard machine learning classification algorithms, using a second layer paradigm to obtain a bet- ter classification accuracy. Experimental results have been obtained us- ing a datafile of cases that contains information about surfaces, extracted from images obtained by the robot. The classification problem consists of recognizing to which of the surfaces belongs a n × n size subimage. The accuracy obtained using the presented new approach statistically improves those obtained using standard machine learning methods. Keywords Supervised Classification, Image Analysis, Image Segmentation, Machine Learning, Stacked Generalization, Classifier Combination 1 Introduction Huge research has been carried out in the field of image segmentation (see [17] for a detailed introduction); more specifically, some authors have dealt with the problem of using color and texture information from images to obtain a good classification of the underlying surface [4]. Supervised image segmentation is a particular kind of supervised classification, in which the objective is to classify each image pixel in order to be able to distinct the different surface segments of the scene in the image. The information obtained from the segmentation process can be used as a first step towards a high-level processing of visual information [2, 16]. But visual information is complex and it is hard to extract useful data in real- time in order to, for example, navigate in the robotics area [8]. Some attempts use optic flow techniques [5] to navigate in semi-structured environments, under the