Modelling plant leaf shape for plant recognition

It is an accepted fact that the challenge in applying image analysis to agricultural processes primarily lies in designing robust image analysis algorithms capable of coping with the combination of an unstructured environment and the inherent variability of biological objects. An example of such a challenging agricultural process could be the discrimination between crop and weed plants from images for selective application of herbicides. This paper describes the initial experiments in modelling the shape variations of plant leaves using image analysis. The shape variability of single leaves from a collection of 385 images of seven different plants is investigated using so called Point Distribution Models (PDMs) (Cootes et al., 1992). The PDM is a general, linear, non-rigid shape model which - for a particular leaf type is computed here by evaluating the statistics of homologous boundary points over a training set of leaf shapes. Even for highly complex leaf shapes, this analysis results in a deformable model governed by a small number of linearly independent parameters. Each of the parameters describes one mode of variation present in the collection of shapes, e.g., the width of the leaf or the length of leaf lobes. It is found that, say, 90% of the total variation of leaves with a relatively simple shape can be described by only 4 parameters and, for complex shaped leaves, only 11 parameters are needed. Typically, the most dominant parameter accounts for more than 40% of the total variation in the PDM. Some problems encountered in applying the PDM - incorrect description of shape variation, shape sets having extreme variation and the automatic extraction of homologous points - are also addressed. In particular, the first results in automatically extracting homologous points using a variation of the curvature scale space concept are presented.