AADB

现实世界的应用程序可以受益于自动生成细粒度照片美学排名的能力。然而，以前的图像美学分析方法主要集中在将图像粗略的二元分类为高审美或低审美类别。在这项工作中，我们建议学习一个深度卷积神经网络来对照片美学进行排名，其中照片美学的相对排名直接在损失函数中建模。我们的模型结合了有意义的照片属性和图像内容信息的联合学习，可以帮助规范复杂的照片美学评级问题。
为了训练和分析这个模型，我们组装了一个新的美学和属性数据库（AADB），其中包含由多个人类评分者分配给每个图像的美学分数和有意义的属性。跨图像记录匿名评估者身份，允许我们在计算训练图像对的排名损失时使用新的采样策略利用评估者内部一致性。我们表明，面对具有不同审美品味的个体对图像审美的主观判断，所提出的采样策略非常有效且稳健。实验表明，我们的统一模型可以生成更符合人类评级的审美排名。为了进一步验证我们的模型，我们表明，通过简单地对估计的审美分数进行阈值化，我们能够在现有的 AVA 数据集基准上实现最先进的分类性能。

Real-world applications can benefit from the capability of automatically generating fine-grained photo aesthetic rankings. However, prior image aesthetic analysis methods primarily focused on coarsely binary-classifying images into high-aesthetic or low-aesthetic categories. In this work, we propose to learn a deep convolutional neural network to rank photo aesthetics, where the relative ranking of photo aesthetics is directly modeled in the loss function. Our model combines joint learning of meaningful photo attributes and image content information, which can help regularize the complex photo aesthetic rating problem.
To train and analyze our model, we assemble a new Aesthetic and Attribute Database (AADB), which contains aesthetic scores and meaningful attributes assigned to each image by multiple human raters. We record anonymous rater identities across images, allowing us to leverage rater internal consistency with a novel sampling strategy when calculating ranking losses for training image pairs. We show that the proposed sampling strategy is effective and robust in the face of subjective judgments of image aesthetics made by individuals with diverse aesthetic tastes. Experiments demonstrate that our unified model can generate aesthetic rankings that better align with human ratings. To further validate our model, we show that by simply thresholding the estimated aesthetic scores, we are able to achieve state-of-the-art classification performance on the existing AVA dataset benchmark.

搜集汇总

背景与挑战

背景概述

AADB数据集是一个专注于照片美学质量评估的图像数据集，由加州大学尔湾分校于2016年发布。该数据集通过深度卷积神经网络结合属性与内容信息，实现细粒度的美学排名，而非传统的二元分类，并包含多个人类评分者分配的美学分数和属性，以提高模型与人类评级的一致性。