基于排序的图像美学质量评估

发布时间：2018-09-05 07:25

【摘要】：图像美学质量评估是计算机视觉领域热点研究问题,并且极具应用前景,可与众多实际应用深度结合。众多相关工作将其定义为分类问题,使用"高质量"、"低质量"标签描述图片美学质量,结合美学特征和分类标签学习分类模型。由于分类标签为粗粒度表达形式,实用性低,因此本文提出研究细粒度的美学质量相对排序。尽管相关工作提出的分类模型以及回归模型亦可解决美学排序任务,但是其仅利用图片绝对美学质量,而非相对排序关系,因而其排序性能较差。一些工作提出基于列排序的方案进行排序预测,然而其训练过程中:1)利用所有可能的排序关系,但是内容差异大的图片间比较却并不合理;2)提取预定义特征为美学特征,然而此特征无法充分描述图像美学属性。针对以上问题,本文首先提出基于图片对的美学排序方案,在训练集中构造图片对反映排序关系,以此为样本训练排序预测模型,且为去除噪声提升质量,应用合理图片对筛选策略对样本进过滤。为进一步提升样本质量,本文提出基于图像检索的图片对构造策略,首先搭建一基于内容的图像检索系统,利用图像检索返回视觉内容相似图片,在此范围内,依据特定生成准则,将查询图片与相似图片构成图片对。为摆脱预定义美学特征的限制,本文提出基于深度学习的美学排序方案,搭建双通道卷积神经网络并设计相应排序损失层,以图片对为输入,以排序关系为目标优化网络权值参数。测试过程中,排序模型计算并输出图片的相对美学排序得分,并以此进行排序,而得分的绝对值大小并无意义。为验证所提出方案的有效性,本文在两个大型公开数据集CUHKPQ和AVA进行美学排序实验,并与其他各方案进行对比,而实验结果则印证了所提出方案在美学排序任务上的优越性。
[Abstract]:Image aesthetic quality assessment is a hot research issue in computer vision field, and has great application prospect, which can be combined with many practical applications. Many related works define it as a classification problem, and use the "high quality" and "low quality" labels to describe the aesthetic quality of pictures, and combine the aesthetic features with the classification label learning classification model. Because the classification label is coarse-grained expression form and low practicability, this paper puts forward the relative ranking of aesthetic quality of fine-grained research. Although the classification model and regression model proposed by related work can also solve the task of aesthetic ranking, their sorting performance is poor because they only use the absolute aesthetic quality of pictures rather than the relative sort relation. Some work put forward the method of ranking prediction based on column ranking. However, in the process of training, the ratio 1) uses all possible sort relationships, but the comparison between pictures with great difference in content is not reasonable. 2) extracting predefined features as aesthetic features. However, this feature can not fully describe the aesthetic attributes of images. Aiming at the above problems, this paper first puts forward an aesthetic sorting scheme based on picture pair, constructs the image pair to reflect the sort relation in the training set, and takes it as the prediction model of sample training ranking, and improves the quality of the training in order to remove the noise. A reasonable picture pair screening strategy was used to filter the samples. In order to further improve the quality of samples, this paper proposes an image pair construction strategy based on image retrieval. Firstly, a content-based image retrieval system is built. According to the specific generation criteria, the query picture and the similar picture constitute a picture pair. In order to get rid of the limitation of predefined aesthetic characteristics, this paper proposes an aesthetic ranking scheme based on deep learning, builds a two-channel convolution neural network and designs the corresponding ranking loss layer, taking the image pair as the input. The network weight parameters are optimized with the order relation as the objective. In the process of testing, the sorting model calculates and outputs the relative aesthetic ranking score of the picture, and then sorts it, but the absolute value of the score is meaningless. In order to verify the validity of the proposed scheme, this paper carries out aesthetic sorting experiments in two large open datasets, CUHKPQ and AVA, and compares them with other schemes. The experimental results confirm the superiority of the proposed scheme in the task of aesthetic ranking.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】