Application of Neural Networks for Seed Germination Assessment MIROLYUB MLADENOV, MARTIN DEJANOV Department of Automatics, Information and Control Engineering University of Rouse 8 Studentska Str., 7017 Rousse BULGARIA mladenov@ru.acad.bg , mdejanov@ru.acad.bg , http://www.ru.acad.bg Abstract: This paper is focused on some key problems, related to the development of a new technology for seed germination assessment using computer vision. An approach for germinated seed image segmentation based on artificial neural networks is proposed. Тhe possibility to use standard neural networks for seed, germ and roots zones extraction is analyzed. A modified RBFN is developed. The accuracy of realized procedures is evaluated. Key-Words: neural networks, computer vision, image analysis, germination, segmentation, color and texture models. 1 Introduction The seeds are one of the important plant agricultural products. The quality and quantity of the plant crops depend on the seed quality. The seed quality is evaluated on the basis of specific sowing properties like purity, germination, humidity, infections, vitality, vigor, authenticity, etc. The diversity of the seed quality factors, as well as the difficulties in their quantitative/qualitative assessment, makes the quality assessment procedures materially difficult. A big part of existing seed quality assessment technologies and tools is primitive and does not change for years. The key role in these technologies has the expert - his knowledge, experience and possibilities for result interpretation. This defines the assessment nature – subjective, slow and expensive. There is a tendency for development of new automatic technologies for seed quality assessment. Computer vision systems are in the basis of these technologies. This is not accidentally. A big part of seed quality factors is evaluated by the expert on the grounds of different visible features. Using image processing methods can be detected different seed quality features like seed shape [1,2,11,12], morphological [3,6,12], color [4,7,14,17] and texture features [8,18], the presence of different injuries [3,12,13] and diseases [18] and can be evaluated different seed sowing properties like purity [1,12,13,17], germination [12,19], infections [18], authenticity [2,5,6,15], etc. The neural networks are widely used in these analyses [2,3,5,9,11,14,16]. They give a possibility to develop more powerful, effective and universal tools for solving different classification problems, related to the seed sowing properties assessment. This paper includes results from an investigation, concerning the usage of color and texture models, as well as neural networks for seed germination assessment based on a computer vision system. 2 Problem Formulation The seed germination is one of the main seed sowing properties. The complete germination assessment by a computer vision system is a difficult and complex problem. The development of procedures for germination assessment is related to the following main problems: germinated seed zone extraction; germinated seed zone segmentation and detection of seed, germ and roots zones; determination of the segment geometrical characteristics; seed classification in quality groups according to the normative requirements. The investigation aims to determine which image features (color, texture) and which image processing tools give an effective solution for seed segment zone extraction. 3 Problem Solution 3.1 Image homogenous zones extraction 3.1.1 Color and texture models The results in related papers [7,8,11,14] show, that an effective result concerning the seed quality assessment can be obtained using image zone segmentation based on colour and texture features. For homogenous zones extraction are used 4 colour models (HSI (Hue Saturation Intensity), XYZ, NTSC(National Television Systems Committee), YCbCr) and the following new texture models: 9th WSEAS International Conference on NEURAL NETWORKS (NN’08), Sofia, Bulgaria, May 2-4, 2008 ISBN: 978-960-6766-56-5 67 ISSN: 1790-5109