毕业生知识库

内容相关的图像缩放技术可以解决图像在尺寸和纵横比发生改变时的内容形变问题，可有效保护图像中主体内容的形状，因此在平面设计、图片网站内容浏览和互联网视频封面制作等方面有着重要应用。然而，由于内容、构图和目标尺寸的不同，任一图像缩放算法在对不同的图像进行缩放时均可能产生不同质量的缩放结果。换言之，不同图像对内容相关缩放操作存在不同的适应度。立足于此，本工作首次提出可缩放度这一图像高层属性，用来描述图像对内容相关缩放算法的适应程度，使用相关视觉属性构建基于卷积神经网络的可缩放度学习与预测方法，并基于可缩放度设计相关应用。本论文的主要研究工作和贡献总结如下：

研究图像本身对内容相关缩放操作的适应程度。从图像高层属性分析角度定义了“可缩放度”，并分析“可缩放度”与其他视觉属性的关联关系。为了学习和分析可缩放度，本文构建了一个涵盖类别更广、数量更多、限制更少的基准图像库（Retargeting and Attributes Database，RADB），该图像库包含专家定义的视觉属性信息和专家评估的每个图像的可缩放分数。

构建基于卷积神经网络的图像可缩放度的学习模型。本文通过对模型的分析与优化，提出了能够描述用户视觉倾向性与图像可缩放度之间关联关系的模型和预测方法——基于深度学习的多任务图像属性学习框架。该学习框架允许模型共享不同视觉属性之间的信息，利用视觉属性共享特征挖掘和预测图像可缩放度。

设计可缩放度相关应用。为切实解决缩放算法在实际应用中的落地问题，本文结合上述学习模型，提出了无结果参照的最优图像缩放算法选择策略；并基于可缩放度，提出平面构图设计时图像素材优化排版的优化方法。

综上所述，本文围绕图像的“可缩放度”这个问题，以互联网海量图像为研究对象，用深度学习的方法对图像“可缩放度”进行分析与挖掘，实现了针对特定图像的可缩放度预测和缩放算法智能选择。

Content-aware image retargeting (CAIR) addresses the increasing demand to display image contents on devices of different resolutions and aspect ratios while preserving its visually important content and avoiding noticeable artifacts. As a result,content-aware image retargeting (CAIR) is widely used in graphic design, photo websites browsing, online-video cover making and so on. Even if the state-of-the-art image retargeting techniques can successfully handle more and more images, it is often not clear beforehand whether a specific image would be successfully retargeted.

In this dissertation, we firstly introduce the notion of image retargetability as a high-level attribute to describes how well a particular image can be handed by content-aware image retargeting(CAIR). Furthermore, we propose a method to learn and predict the retargetability, taking advantage of convolutional neural network and related visual attributes. To further validate our model,we show applications of image retargetability. The issues and contributions being addressed include:

We study the adaptation of the image itself to content-aware image retargeting(CAIR).We proposed retargetability as a novel image property and developed a computational predictor for it. At the same time, we analyze the correlation between retargetability and attributes.To learn and predict retargetability, we have assembled a new retargeting and attributes database (RADB) which contains retargetability scores and meaningful attributes assigned to each image by an expert rater.

We build and modify the learning model of retargetability based on convolutional neural network. As to the automatic retargetability scoring, we demonstrate the multi-task learning approach by jointly learning visual attributes from deep features and feature sharing for retargetability. This approach allows the model to share information among different visual attributes and by making use of the shared feature, the model mines and predicts retargetability  from an image.

We design applications related with retargetability. We propose the strategy to choose the "best" image retargeting algorithm without result-images. We show how to utilize image retargetability in applications of retargeting method assessment, photo collage and photo album cover generation. And we propose the method to optimize the layout of image material in the graphic design.