Abstract
Evaluation of soybean plant shape in breeding is empirical and based on visual judgment, making it unstable and inefficient and pointing up the need for a quantitative alternative. Previous studies successfully applied evaluation by linear discriminant function, fuzzy logic or neural network, but these models required definition and selection of important features for judging shape. We developed a method based on a multilayer perceptron (MLP) with direct image input of binary soybean images which does not require any shape features. An MLP is a kind of neural networks, and can exhibit good performance in pattern recognition. A neural network is composed of units being simple processors, and connections between the units carrying numeric data from one unit to another. Units of an MLP are arranged on the layers, and connected each other between the adjoined layers. We used 326 soybean plant images judged either”Good” ”Fair” or ”Poor” by expert soybean breeders. The images were divided into supervisor and test data sets. We studied 175 different MLP structures, varying the number of layers, units and connections. After training each MLP with the supervisor data set, we evaluated matches between MLP output and breeder judgment with the test data set. The MLP with three layers, 8×8 input units, 16 hidden units and three output units proved to be the superior structure. Although performance in judgment was no higher than that of previous ones, our method has the decided advantage of not requiring definition and extraction of the shape features and may be applicable to other crops. We should note that the MLP structure is too complicated for us to understand the n]anner of breeders' empirical judgment through this model; that is, the MLP is almost a black box for us for the time being.
