Towards Automatic Segmentation of Serial High-Resolution Images

Abstract

Developing barley grains are to be visualised by a 4-D model, in which spatiotemporal experimental data can be integrated. The most crucial task lies in the automation of the extensive segmentation procedure. Because of constraints like incomplete a-priori expert knowledge and the complexity of this specific segmentation task, learning techniques like Artificial Neural Networks (ANN) yield promising solutions. In this work we present our first good segmentation results. Two different supervised trained ANN classifiers were applied, on one hand, the well-established borderline-learning Multiple-Layer Perceptron (MLP) and on the other hand, the prototype-based Supervised Relevance Neural Gas (SRNG). While so far segmentation was mainly achieved using manual tools, now almost automatic segmentation becomes more feasible.