Discrimination of soybean leaflet shape by neural networks with image input

It is necessary to correctly evaluate intra- and inter-specific variations for the efficient collection and preservation of genetic resources, and leaf shape is one of the important characteristics to be evaluated. It has been thought that a more consistent and quantitative method should be introduced to aid in the processes of practical discrimination. Several researchers have suggested leaf shape evaluation methods using shape features, and these methods have shown good results. The shape features selected in these methods have differed from one method to another, and new shape features must be redefined when these methods are applied to new cases. The processes for defining and extracting shape features are ad hoc. We, therefore, have attempted to develop a generalized model that requires neither the definition nor extraction of any shape features; the method uses neural networks into which leaf shape images are input. In this study, we applied a Hopfield model and a simple perceptron to the varietal discrimination of individual leaflet shapes of 364 soybean leaflets of 38 varieties. In the examination of up to ten varieties, the discriminant error of the neural networks with image input was satisfactorily low even under cross validation. We, therefore, concluded that this model works quite well for quantitative varietal discrimination in the case of soybean leaflets. The advantage of requiring neither the definition nor extraction of any shape features makes us expect that this model will be widely applicable to other cases, and we will attempt to verify this applicability.