基于卷积神经网络和SVM的中国画情感分类

Chinese Painting Emotion Classification Based on Convolution Neural Network and SVM

图像情感是指计算机识别数字图像所表达内容引起人的情感反应,根据不同的情感反应,可以对不同的图像进行分类.在信息量急剧增长的今天,图像情感分类有助于图像的标注和检索,蕴藏着很大的社会和商业价值.不同于西洋画的"以形写形",中国画有着自己明显的特征:传统的国画不讲焦点透视,不强调自然界对于物体的光色变化,不拘泥于物体外表的肖似,而多强调抒发作者的主观情趣.这比弥合一般的低层特征和人类情感高层语义之间的鸿沟的难度更大.基于卷积神经网络因为其具有结构简单、适应性强、训练参数少、连接点多等特点,可以直接输入原始图像,能够避免对图像进行复杂的前期预处理.相比传统图像特征提取方法,卷积神经网络具有明显的优势.本文的目的是利用卷积神经网络发掘低层特征和情感语义之间的联系,提取国画图像特征,对得到的特征进行PCA降维、归一化等操作后,利用支持向量机(SVM)分类器进行情感分类.

Image emotions are human emotional responses caused by the contents of digital images. Computers are able to classify different images according to different human emotional responses. With the rapid growth of the amount of informa-tion, image emotion classification will contribute to the image annotation and search producing great social and commercial value. Chinese paintings have obvious characteristics ： traditional Chinese paintings do not focus on the perspective, and do not emphasize the light color changes of objects in nature,and do not rigidly adhere to the appearance of objects. They more focus on the expression of authors＇ subjective consciousness making it harder to bridge the semantic gap between general low-level features and human emotions. The structure of convolutional neural network （ CNN） is simple ,yet its adaptability is strong. CNN also has less training parameters and more junctions,and are able to read images directly without preprocessing images complexly. It has a huge advantage over traditional image-processing method. This paper aims to explore the rela-tionships between low-level features and emotional semantics by CNN, and extract the features of Chinese paintings and process the features by PCA and normalization. Finally we classify the features by SVM.