基于多尺度特征提取网络的图像美学量化评分方法

提出了一种基于多尺度特征提取网络的图像美学客观量化评分方法，该模型主要由多个多尺度特征提取单元级联组成，每个单元包含由3个不同卷积核组成的特征提取层、融合层和映射层。特征提取层通过联合图像的全局视图和局部视图组成网络输入端，在输出端以EMD函数为损失函数，输出分布为1~10分的概率密度质量函数，并以分布均值作为图像美学量化值。实验证明，本文方法具有可行性和有效性，解决了传统方法只进行美感二进制等级分类的问题，给出了（模拟人类思维对）图像的客观量化评分；同时在AVA数据集上获得了优于几种主流算法的分类准确度。

Quantitative scoring method of image aesthetics based on multi-scale feature extraction network

An objective quantitative scoring method of image aesthetics is proposed based on multi-scale feature extraction network.The proposed model mainly comprises several multi-scale feature extraction units,each of which includes a set of feature extraction layers with different convolution kernels,a fusion layer and a mapping layer.The feature extraction layer combines the global view and the local view of the image to form the input of the network.The EMD function is used as the loss function in the softmax layer.The output is a probability density mass function from 1 to 10,and the mean is used as objective qualitative score of picture quality.Experiments show that the proposed algorithm is feasible and effective,in particular,it solves the problem that the traditional method obtains only the binary classification of aesthetic,and the classification accuracy of AVA dataset is better than that of several mainstream algorithms.