杨冰 陈浩月 王小华

摘 要：绘画图像分类为绘画管理与使用提供了便利。传统图像分类大多依靠人工提取形状、颜色等特征，由于绘画图像分类需要更专业的知识背景，从而使人工提取特征的过程繁琐且复杂。基于此，提出一种基于卷积神经网络的中国绘画分类方法，并在此基础上结合SoftSign与ReLU两种激活函数的优点，构造一种新的激活函数。实验结果表明，基于改进后激活函数构造的卷积神经网络，可以有效提高分类准确率。

关键词：深度学习;卷积神经网络;中国绘画;激活函数;图像分类

DOI：10. 11907/rjdk. 181736

中图分类号：TP301文献标识码：A文章编号：1672-7800（2019）001-0005-04

Abstract：The classification of painting images facilitates the management and use of paintings. Different from traditional image classification， features such as artificial extraction of shapes and colors are required. Classification of painting images requires a more professional knowledge background， which also makes the process of manually extracting features increasingly complicated. Based on this， a Chinese painting classification method based on convolutional neural network is proposed. Based on this， it combines the advantages of two activation functions including SoftSign and ReLU to construct a new activation function. Experimental results show that the convolutional neural network constructed based on the improved activation function can effectively improve the classification accuracy.

Key Words： deep learning; convolution neural network; Chinese painting; activation function; image classification

0 引言

随着数字化图像的发展，图像分类成为图像领域的研究热点之一。作为中国传统文化的重要组成部分，对中国绘画分类[1-2]的研究有助于更好地继承与发扬传统文化。中国绘画历史源远流长，流派与艺术风格众多，且中国画以写意为主，与自然状态下的图像相差较大，所涵盖的内容也更加抽象，所以在特征提取方面需要更多专业知识。传统图像分类方法大多是基于浅层结构的学习算法，虽然可以提取一定图像特征，但在某些特征提取过程中容易导致特征丢失，且特征提取方法的泛化性较差。因此，中国绘画的图像分类存在诸多障碍。

受Hubel & Wiesel对于猫视觉皮层电生理研究的启发，卷积神经网络由此诞生。Yann Lecun首次将卷积神经网络用于手写数字识别[3-4];Krizhevsky等[5]提出经典的卷积神经网络结构AlexNet，并在图像识别任务上获得重大突破。……