Plant leaves classification based on morphological features and a fuzzy surface selection technique Panagiotis Tzionas Technological Educational Institute of Thessaloniki, Thessaloniki 574 00, Greece. e-mail:ptzionas@teithe.gr Stelios E. Papadakis Technological Educational Institute of Kavala, Kavala 65404 , Greece. e-mail:spap@teikav.edu.gr Dimitris Manolakis Technological Educational Institute of Thessaloniki, Thessaloniki 574 00, Greece. e-mail:dmanol@teithe.gr Abstract— The design and implementation of an artificial vision system that extracts specific geometrical and morphological features from plant leaves is presented in this paper. A subset of significant image features are identified using a novel feature selection approach. This approach reduces the dimensionality of the feature space leading to a simplified classification scheme appropriate for real time classification applications. A feed- forward neural network is employed to perform the main classi- fication task. The proposed system exhibits size and orientation invariance with respect to the samples and it can operate successfully even with leaves samples that are deformed due to drought or due a number of holes drilled in them. A considerably high classification ratio of 99% was achieved, even for the classification of deformed leaves. Index Terms - Feature selection, classification, fuzzy sur- face, neural networks, image processing. I. I NTRODUCTION The classification of plant leaves is a crucial process in botany and in tea, cotton and other industries [1], [2]. More- over, the morphological features of leaves are used for plant classification or in the early diagnosis of certain plant diseases [3]. This paper presents the design and implementation of an artificial vision system capable of extracting geometrical and morphological features from plant leaves. Initially, leaves taken from plants in the native environment and surroundings were collected and used as samples for testing the proposed system. Later, additional samples originating from diverse environments were used for classification. The proposed system consists of: a) an artificial vision system (frame grabber and camera) b) a combination of image processing algorithms implemented in LabView [4] and c) a feed-forward neural network based classifier implemented in MatLab [5]. The image processing part is responsible for image capture and image pre-processing in order to obtain normalized fea- tures [6], [7] and for determining some critical geometrical characteristics. The study of such morphological features has been extensively used in the literature [8], [9], [10]. How- ever, the plethora of geometrical and morphological features makes it impossible to use all available features in a certain classification problem, especially in real-time applications and, thus, some selection technique is necessary. A novel, fast and consistent approach for calculating the importance of each subset of features, that together are assumed to influence the classifier output of the image processing system the most, from a set of candidates features, is presented in this work. More specifically, a fuzzy surface technique [11] is used for building fast a coarse model of the system from a subset of the initial candidate features. A neural network is then trained with the selected morphological features and classifies the feature space to appropriate categories. Neural networks have been also used extensively for classification [12], [13]. In order to develop an efficient classifier two fundamental and contradicting modeling principles should be satisfied: a) maximization of the identification/generalization capabilities and b) minimization of the architectural complexity. Since the neural network complexity increases exponentially with the number of inputs, an input selection technique is re- quired to identify the significant features from the plethora of candidate geometrical and morphological features. This data pre-processing task is carried out using an entirely different modeling technique to the neural network classifier modeling. This technique [11], based on fuzzy surfaces, emphasizes on the maximization of learning-generalization capabilities and ignores the complexity of the architecture for the sake of modeling celerity. Furthermore, the proposed system is also capable of auto- mated image capture (for ’production line’ operation), count- ing the number of the leaves per category, counting of the holes that may be present on the leaves surface (due to diseases, malformations etc.) and morphological classification of these holes. This paper is organized as follows: Section II provides the image processing algorithms for morphological feature extrac- tion. Section III presents the feature selection problem and a novel fuzzy surface approach adopted for the determination of the significant subset of features. Having decided on the optimal subset of significant features, the optimal structure of the reduced dimensionality neural network classifier is presented in section IV. Section V provides a description and evaluation of the overall image processing system. Finally, the advantages of the proposed system are highlighted in the conclusions section. II. MORPHOLOGICAL FEATURE EXTRACTION Several morphological and geometrical features are ex- tracted from the leaves using the proposed image processing