Plant Shape Discrimination of Several Taxa without Shape Feature Extraction Using Neural Networks with Image Input.

Oide Mari, Ninomiya Seishi

doi:10.1270/jsbbs.50.189

The importance of developing a quantitative shape discriminant method has been widely emphasized to obtain a consistent and objective evaluation. In fact, several methods for shape discrimination have been proposed. However, each of the methods can only be applied to a specific material, requiring specific shape features. Considering the number of shape discrimination problems in agriculture, such as genetic resource management, a single model that can be generally applied to different discrimination problems is highly desirable. A shape discriminant model based on neural networks with image input that did not require any shape features has been recently proposed, and it was shown that this model could be applied to soybean leaflet shape discrimination. This model was expected to be highly suitable for various materials, because it did not require any shape features. In this study, we examined the applicability of this model to the discrimination of several plant materials, by changing the training conditions of the neural networks to find their optimal combination. The plant materials were as follows: maple leaves, Tartary buckwheat kernels, pear fruits, mulberry leaves, and leaf-mustard leaves. The discriminant error rates, which were evaluated by cross-validation examination, differed depending on the materials, the number of training classes, and the training conditions. For example, the cross-validation error rate was 0.202 in the case of maple leaf shape with 10 training classes under the optimal training condition. The optimal training condition was common to all the materials examined in this study. Considering that this model did not require any shape features, its error rate was acceptable. This model can be widely and easily applied even for the evaluation of rather complex shapes such as lobed leaves. We concluded that this model would be highly suitable when various shapes of phytoorgans in many species have to be discriminated, such as in the case of genetic resource management.

Plant Shape Discrimination of Several Taxa without Shape Feature Extraction Using Neural Networks with Image Input.

Search this article

Description

Journal

Citations (2)*help

References(20)*help

Keywords

Details 詳細情報について

Export

Report a problem