基于信息熵的风格绘画分类研究

Artistic Paintings Classification Based on Information Entropy

针对艺术风格绘画分类算法中存在的精度和效率不高等问题,提出一种基于信息熵的艺术风格绘画分类算法。首先选取西方漫画、素描、油画、水彩画,以及国内烙画、水墨画、壁画具有代表性的7种艺术绘画风格作为研究对象,对图像进行去噪、归一化等预处理。其次,提取绘画艺术作品风格特征,分别求取图像的颜色熵、分块熵、轮廓熵,并合并构成不同绘画风格的信息熵。信息熵求取时,将色彩空间转换为Lab颜色空间,通过a、b通道颜色值及加权函数获得图像的颜色熵;通过对艺术图像分块求取信息熵,求取分块的信息熵均值获得分块熵;通过Contourlet变换,求取艺术图像的轮廓信息,获得轮廓熵。接着,合并提取的颜色熵、分块熵、轮廓熵,利用支持向量机(SVM)对艺术风格图像学习训练,获得艺术绘画风格的分类模型;最后,提取待识别绘画风格样本的熵特征,通过SVM分类识别获得最终的分类结果。该方法具有特征维数少、运算速度快、尺度不变性等优势,实验结果表明,其能提高不同绘画风格的分类精度和效率。

Aiming at the improvement of the accuracy and efficiency of artistic paintings classification algorithm, this paper puts forward a style painting classification algorithm based on information entropy. Firstly, seven representative painting styles, including cartoons, sketches, oil paintings, watercolor paintings, and art painting styles of Chinese pyrography, ink painting and murals, were selected as the research objectives, and the images are pre-processed by denoising and normalization. Secondly, we extracted the style features of painting images and obtain the color entropy, block entropy and contour entropy respectively. Then, the algorithm combined the information entropy of different input painting styles. During the calculation of the information entropy, the color space was transformed from RGB to LAB, and the image color entropy was obtained from a and b channel values and weighting functions. By dividing the artistic images into blocks, we calculated the average entropy of all the blocks to obtain block entropy. Contourlet transform was used to obtain the contour information of artistic images, and we obtained contour entropy. After that, color entropy, block entropy and contour entropy were merged and extracted, and support vector machine(SVM) was applied to train the artistic style image to obtain the classification model of artistic paintings. Finally, we extracted the entropy characteristics of the samples to be identified, and obtained the final classification results by SVM. The method proposed has the advantages of less feature dimension, fast operation and scale invariance. The experimental results show that the proposed method can improve the classification accuracy and efficiency of different painting styles.