可计算的图像美学分类与评价系统研究

发布时间：2018-03-19 12:29

本文选题：图像美学分析　切入点：美感分类　出处：《华南理工大学》2013年硕士论文　论文类型：学位论文

【摘要】：可计算图像美学的研究目的是希望计算机能够模拟人类的视觉系统与审美思维对图像进行美学价值的判断。其研究结果可以应用到融合主观感知的基于语义的图像检索、图像美学质量评估、摄影的美学预测与修正、艺术作品风格分析、人机交互，以及设计、摄影、广告等领域。本文以数字图像和摄影照片为研究对象，探索采用计算机进行图像美学自动分类与评价的可计算方法，，通过对图像划分整体区域和关键区域，提取有效的低层视觉特征和高层美学特征，并采用机器学习的方法建立图像美感等级评估分类器和美学分数预测模型，实现了机器自动评估图像的高、低美感并预测图像的美学分数。围绕可计算图像美学的分类评价研究，本文的工作包括： 1、基于人类视觉心理学美感研究，建立了包含人类审美评价信息的美学图库。 2、设计了一种结合分水岭分割算法与图像梯度特征的主体区域提取方法，提取主体区域作为图像的关键区域。 3、完成了基于美学的图像特征提取，包括图像低层特征、高层美学特征及区域特征。 4、对特征数据进行机器训练学习，采用AdaBoos（tAdaptive Boosting，自适应增强算法）分类算法建立图像美感等级评估分类器，并通过SVR（Support Vector Regression,支持向量回归）算法建立图像美学分数预测模型。 5、实现了可计算的图像美学分类评价系统。本文对大量含有美感信息的图像进行测试，实现高低美感分类的平均分类准确率为77.4%，而回归预测相关性为0.795、均方根误差为0.244。实验证明了本文算法的有效性，系统美感的结果与人类对图像审美感知结果高度相关。
[Abstract]:The purpose of the research on computable image aesthetics is to hope that the computer can simulate human visual system and aesthetic thinking to judge the aesthetic value of images. The research results can be applied to semantic image retrieval based on subjective perception. Image Aesthetics quality Assessment, Photography Aesthetics Prediction and Correction, Art style Analysis, Human-Computer interaction, Design, Photography, Advertising, etc. In this paper, we take digital images and photographic photographs as research objects, and explore the computable method of automatic classification and evaluation of image aesthetics by using computer. By dividing the image into the whole region and the key area, Extraction of effective low-level visual features and high-level aesthetic features, and the establishment of image aesthetic evaluation classifier and aesthetic score prediction model by machine learning method, which can automatically evaluate the image height. Low aesthetic sense and predict the aesthetic score of the image. Based on the research of classification and evaluation of computable image aesthetics, the work of this paper includes:. 1. Based on the study of aesthetic sense of human visual psychology, an aesthetic library containing human aesthetic evaluation information is established. 2. A main region extraction method combining watershed segmentation algorithm with image gradient feature is designed. The main body region is extracted as the key region of the image. Thirdly, the extraction of image features based on aesthetics is completed, including image low-level features, high-level aesthetic features and regional features. 4. The feature data is trained by machine, and an image aesthetic evaluation classifier is established by using AdaBoos(tAdaptive boost (adaptive enhancement algorithm) classification algorithm, and an image aesthetic score prediction model is established by SVR(Support Vector regression (support vector regression) algorithm. 5. A computable image aesthetic classification and evaluation system is implemented. In this paper, a large number of images with aesthetic information are tested, and the average classification accuracy is 77.4, the correlation of regression prediction is 0.795, and the root mean square error is 0.244.The experimental results show that the algorithm is effective. The result of systematic aesthetic sense is highly related to the result of human aesthetic perception of image.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】