Digitization for CBIR

The vast majority of searches on images are metadata-based. Although it is efficient to search for text-based metadata, such systems require humans to enter descriptions about each image in the repository. The consistency and completeness of metadata are always big problems.

For the past two decades, researchers have been working on various algorithms to retrieve desired images from a large collection by analyzing contents of images. This research area is generally called content-based image retrieval (CBIR). The basic idea of CBIR is to extract feature items from an image and make comparison with the target item or search reference. CBIR tools rank the images in the repository and present a few of the closest ones. The most commonly used feature items for comparison are color, texture, and shape, all of which are in pixel spaces. Stepping further, other researchers have tried to advance the searches into semantic spaces.

Color is the lowest level of physical characteristics that can be extracted from an image. It is relatively robust to measure and analyze. Texture refers to visual patterns in images. All surfaces, such as trees or clouds, can be modeled with texture. It contains essential information about the structure of surfaces and their relationship to surrounding environment. Shape refers to the boundary and region of a geometric object.

If digitization for CBIR is a major aim, post-scan processing is important in order to facilitate the feature extractions of color, texture, and shape. It is thus appropriate and desired to make the following adjustments which are recommended by NARA “Technical Guidelines for Digitizing Archival Materials for Electronic Access.”

1. Color and gamma correction
2. Tonal scale adaptation
3. Texture filtering to compensate for variations in originals
4. Sharpening to match appearance of the original

In addition to the technical considerations for scanning and image processing, carefully selecting materials for digitization is critical because the execution of CBIR is very time consuming as of today’s technologies.

Advertisements

2 responses to “Digitization for CBIR

  1. How does google image search work when you search using an image? That said, and also with the caveat that i have to keep a picture of google like that one in the beginning of memento “do not trust this man” because I want to so much, what are the options for taking something like the CBIR, and associate it with the other data that resides at google or thereabouts to harvest metadata about the object in the image?

  2. I believe Google uses text metadata rather than CBIR. We enter keywords, rather than a reference image or a drawing sketch for search. In the future, text-based metadata together with CBIR will push image discovery and access to an advanced level. However, CBIR has still a long way to go, as I understand from the literature I read.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s