基于神经网络的批量定制服装号型分类研究

Investigation on Mass Customization Clothing Shape Classification by Artificial Neural Network

为提高批量定制服装号型的分类效率,运用神经网络方法,以男衬衫为研究对象,用6个测量项目(身高、胸围、腰围、颈围、肩宽和全臂长)为分类变量,对686名男性人体号型进行K-means聚类分析,并将聚类结果作为样本,对神经网络进行训练和测试。以混淆矩阵为指标研究了不同网络结构、训练算法及传递函数的分类效果。研究表明,分类效果随训练算法、网络结构不同存在明显差异,其中标准BP算法分类效果最差,弹性BP算法分类效果最好,且分类效果随隐层神经元数量的增加而提高,隐层和输出层传递函数均为logsig时,分类效果最好。

In order to accelerate shape classification efficiency of mass customization clothing, artificial neural network method was applied. Take man shirt for example, 686 male body data was classified by K-means algorithms with 6 variables（height, bust circumference, waist circumference, neck circumference, shoulder width and arm length） to generate sample for training and testing neural network. Then, classification accuracy rate of network was studied by confusion matrix with different structures, training algorithms and transfer function. The results revealed obvious difference of classification accuracy rate existed in network with different training algorithms and structure. The lowest classification accuracy rate came from standard BP algorithms, while the highest one came from resilient BP algorithms. The classification accuracy rate increased with the increase of the neuron number of hide-layer. The best classification came from logsig transfer function both used in hide-layer and output-layer.