Neural network analysis of MINERVA scene image benchmark

Abstract

Scene analysis is an important area of research with the aim of identifying objects and their relationships in natural scenes. MINERVA benchmark has been recently introduced in this area for testing different image processing and classification schemes. In this paper we present results on the classification of eight natural objects in the complete set of 448 natural images using neural networks. An exhaustive set of experiments with this benchmark has been conducted using four different segmentation methods and five texture-based feature extraction methods. The results in this paper show the performance of a neural network classifier on a tenfold cross-validation task. On the basis of the results produced, we are able to rank how well different image segmentation algorithms are suited to the task of region of interest identification in these images, and we also see how well texture extraction algorithms rank on the basis of classification results. Scene analysisOutdoor scene analysis is a complex problem. A number of different approaches have been used for recognising different objects in such scenes. In early experiments on scene analysis, simple problems were tackled. For example, Brice and Fennema [7] defined the procedure to interpret simple objects in images such as wedges, cubes, wall and floor. Model-based approaches, such as the one proposed by Brooks [8], have met with some success, provided that objects can be defined with geometric primitives. These approaches have problems with recognising natural objects such as trees, for example, where such primitives are hard to define. Another approach is based on the use of a knowledge-based scheme where hand-coded rules are used for object recognition. These rules describe the characteristic properties of objects of different types. Some examples of such work include SCHEMA vision system by Draper et al. [12], region-based scene analysis by Ohta [34], and VISOR connectionist system for scene analysis [28]. Although these approaches have shown reasonable results, a significant amount of computational effort is required even for simple scenes. More recently, far more complex problems are being solved using texture-based image analysis. The thesis of Becalick [5] and MacKeown [30] provides an excellent review treatment of other studies in this area and summarise the progress made. In particular, Becalick reviews important studies based on knowledge-based approaches in scene analysis, Autonomous Ground Vehicle (AGV) research, and image database and multimedia search.Batlle et al.[4] review a number of studies in the area of scene analysis for urban and natural scene classification. Campani et al. [9], and Parodi and Piccioli [35] have used colour information in the classification of natural objects including road boundaries, road signs, vehicles, buildings and trees. Colour information has also been used by the following authors: Draper et al. [12], and Hanson and Riseman [15] for the discrimination of objects (road scenes) including sky,