基于迁移学习的名家画作数字化识别与分类

Famous artist’s painting digitization recognition and classification based on transfer learning

在数字化过程中,如何将不同时期不同风格的未署名画作快速分类成为一个难题。针对以上问题,提出一种通过迁移学习的卷积神经网络的识别分类方法。ResNet-50神经网络模型在ImageNet数据集上完成预训练,通过迁移学习的方式将特征参数迁移到WikiArt数据集上处理选取的23位画家的400幅画作(总共9200幅)。通过迁移学习ResNet-50和DenseNet-201两个神经网络模型对比,发现基于ImageNet数据集在ResNet-50的迁移学习神经网络模型具有良好的特征提取能力,在WikiArt数据集上top-1识别率最高达到81.6%。

In the process of digitization,how to quickly classify unsigned paintings with different styles in different periods is a problem.Hence,a recognition classification method based on convolutional neural network with transfer learning was proposed.ImageNet datasets were performed using the ResNet-50 network model on pre-training,and the feature parameters were transferred to the WikiArt datasets by transfer learning to process 400 paintings of 23 artists(total 9200 paintings).Comparing two neural network models of transfer learning ResNet-50 and DenseNet-201,It is found that the transfer leaning neural network model based on ImageNet datasets in ResNet-50 has good feature extraction ability.On Wikiart dataset,the recognition rate of top-1 is 81.6%.