Fast content-aware resizing of multi-layer information visualization via adaptive triangulation
Introduction
Content-aware resizing is an adaptive technique in image processing that filters out less important content and retains more important ones. This technique has also become a useful tool for information visualization because the diversity of displays for hardware is increasing. In addition, virtual displays of arbitrary size or aspect ratio require content-aware resizing techniques. Although artists, web designers, and programmers can design several available layouts for different scenarios, the task is time-consuming and costly.
The main objective of this work is to perceptively adjust the data output to any size or aspect ratio of a target display in the context of multi-layered information visualization applications. The basic resizing techniques such as cropping and linear scaling often result in information loss and visual distortions. Cropping (Fig. 1b) is the simplest operation for visualization resizing. Cropping can be used to adapt to different types of displays. However, cropping often removes important content. Linear scaling (Fig. 1c) is another approach. However, distortions often appear when more important content regions have the same scaling rate as less important regions. Missing content and added distortions in visualizations quickly lead to loss of attention, or worse, the complete misinterpretation of the information presented. Hence, it is paramount to adopt a robust approach that not only resizes the content appropriately but also retains the important content in information visualization.
Existing approaches for content-aware resizing mainly focus on natural images such as portraits, landscapes, and buildings. On the other hand, in information visualization, images normally consist of abstract mathematical representations such as vectors, points, lines, icons and geometrical shapes. The important content often represents the main subject in visual content. Fig. 1(e) shows an example of content-aware resizing of visual information produced by our method.
Most of the existing image resizing approaches are not entirely suitable for information visualization. Grid-based methods [1], as shown in Fig. 1(d), have been used for resizing such images as geometric distortions and are easily identified. Pixel-based seam carving, Avidan et al. [2], is normally used for image resizing. This technique cannot be easily extended to account for layout adjustments of geographical scatterplots and social network graphs. In addition, the criterion for significant regions in visualization is different from natural images because their color schemes are different. For example, a blue sky in a natural scene is normally classified as background. Therefore, it would often be considered less important than a person in the foreground. Unlike natural images, a blue region rendered by a visualization system may be regarded as an important region.
Information visualization often consists of multiple information layers. For example, a geographical application would normally contain several layers such as water, continents, and various location markers. If we ignore major regions such as continents in resizing, then the results will suffer from distortions. Multi-layered based resizing is rarely discussed in the previous work either on image resizing or on information visualization resizing. The resizing framework of Wu et al. [3] assumed the information visualization layer is single such as scatter-plot, network, and word cloud. However, often there are many abstract layers in information visualization designs such as a scatter-plot on a map and graph with group shapes. Therefore, it is necessary to revisit the multi-layer approaches to detect and preserve the different layers in information visualization. When the resizing content is complex and the canvas become larger, the time performance is becoming more important. Prior work such as Wu et al. [3] requires adjustment to fast resize the information visualization.
Hence, based on the resizing pipeline of Wu et al. [3], we present a different visualization resizing approach in three aspects. First, we define a visualization-related saliency map. Second, we consider the classes of information to be segregated into multi-layers for visualization. Third, the controlling mesh for resizing in our approach is adaptive so that users can emphasize the content of the visualization with fewer distortions in a shorter period of time. The contributions of our work are:
- 1.
an abstract multi-layer model for the resizing problem of information visualization. Our model can be used to resize the output from a visualization system to automatically match the native aspect ratio of any external target display;
- 2.
a set of criteria called the visual saliency map (or VSM) to describe the features of information visualizations in different saliency layers;
- 3.
a triangle mesh-based energy optimization method to achieve better visual distribution of information features after resizing. We present the results of our experiments on different genres of multi-layered visualizations to demonstrate the performance of our approach.
Section snippets
Related work
In the following subsections, we review the related methods on content-aware resizing, saliency mapping, and adaptive meshing.
Overview of our proposed method
We define information visualization resizing as a saliency detection and geometric deformation problem. The input of our model is a multi-layered rendering. Multi-layers can be viewed as more than one representation in information visualization. In the example of Fig. 1(a), the input includes a geographical map, lines and nodes with different radii. First, we detect the visual saliency through a hybrid saliency model, the VSM, that can generate different saliencies for different layers in
The visual saliency map (VSM)
The proposed saliency-based method, called the visual saliency map (VSM), adaptively indicates the significant regions in a information visualization. For the content-aware resizing, the deformation of each region is dependent on its corresponding saliency. Although the saliency of each region can be assigned by users manually, it is more effective to automatically detect the important regions.
The saliency concept for visualization is different than the one used in natural images in three
Adaptive resizing model
In the following sections, we describe our resizing model in detail. First, we start with the adaptive meshing. We resort to triangular meshes as triangular meshes can be more readily adapted to high-density regions.
Experiments and results
All the experiments in this paper were performed on a computer with Windows 7 OS, an Intel i7 CPU 2.8 GHz and 8 GB RAM. We implemented the algorithm in C++ and used CGAL [25] library to generate the triangulation. Lapack++ and C++ library were employed to solve the large sparse linear system of equations.
We tested our method on several datasets and obtained better results than previous methods. Because we use the adaptive method to generate triangles, better performance can be achieved than
Conclusion
In this paper we present an adaptive triangle-mesh based method for content-aware resizing of information visualizations. We propose a visual saliency detector that follows seven criteria. The detected visual saliency map (VSM) is not only used to generate adaptive meshes but also used to calculate the deformation factor of each triangle. A robust resizing energy function is defined to implement mesh resizing. The experiments show that our method can be used effectively in redesigning
Acknowledgment
The authors would like to acknowledge the partial support of the Hong Kong Research Grants Council Grants, GRF PolyU 5100/12E, IGRF PolyU 152142/15E and Project 4-ZZFF from the Department of Computing, The Hong Kong Polytechnic University.
References (31)
- et al.
Study of visual saliency detection via nonlocal anisotropic diffusion equation
Pattern Recognit.
(2015) - et al.
Triangulations in CGAL
Comput. Geom.
(2002) - et al.
Robust visual tracking via online multiple instance learning with fisher information
Pattern Recognit.
(2015) - et al.
Optimized scale-and-stretch for image resizing
ACM Trans. Graph.
(2008) - et al.
Seam carving for content-aware image resizing
ACM Trans. Graph.
(2007) - et al.
ViSizer: a visualization resizing framework
IEEE Trans. Vis. Comput. Graph.
(2012) - et al.
A comparative study of image retargeting
ACM Trans. Graph.
(2010) - et al.
Improved seam carving for video retargeting
ACM Trans. Graph.
(2008) - et al.
Structure extraction from texture via relative total variation
ACM Trans. Graph.
(2012) - Y. Pritch, E. Kav-Venaki, S. Peleg, Shift-map image editing, in: Proceedings of the IEEE International Conference on...