Image extraction and preprocessing

We use off-the-shelf tools to extract images from the embedding documents. For instance, images in PDF documents can be extracted by Adobe Acrobat image extraction tools. Images contained within HTML document can be extracted by special HTML parsers. Images extracted from PDF are usually in “PNG” format. Web images are typically in GIF format. Based on our observations, the majority of images extracted from PDF documents are stored in raster format and may also contain color information. Typically, humans do not need to see the images in full color in order to determine the class label of an image, though full color certainly helps in understanding the meanings of the images. Thus, we convert all images to gray scale format in order to standardize the input format of our system. Specifically, we convert all images to the Portable GrayMap (PGM) format, a gray scale
image format which is easy to manipulate.


Extracting text and numerical data from 2-D plots
Two-dimensional (2-D) plots represent a quantitative relationship between a dependent variable and an independent variable. Extracting data from 2-D plots and converting them to a machine-processible form will enable users to analyze the data and compare them with other data. Extracting the metadata related to 2-D plots will enable retrieval of plots and corresponding documents and will help in the interpretation of the data. We developed a system for extracting metadata from single-part 2-D plot images, i.e., a single 2-D plot in the 2-D plot image.


Extracting line features

A part feature refers to a part of an image with some special properties, e.g., a circle or a line. Based on our definitions of several non-photographic image classes and our experimental data, we observed correlations of certain objects with corresponding image classes. For example, a two dimensional coordinate system, consisting of two axes, are
commonly seen in 2-D plots; rectangles, ovals and diamonds are common objects in diagrams. Thus, we attempt to design part image features for basic objects in non-photographic images and use them to discriminate different classes of
non-photographic images.