According to Dr. Mat Malladi, CEO, the Internet is evolving
from a text to visual medium. There are over 1 billion images
on net today, and 7 million images are added each day. However,
current Internet search technologies are limited to performing
text and keyword searches. Users typically receive thousands
and even hundreds of thousands of results to these types of
searches, an unwieldy amount of information. eVision's visual
recognition technology provides a means for conducting large-scale
searches on rich media content.
According to eVision, although some visual searches do exist
today, these competing visual search schemes rely on manual
annotation of the images and videos, and when a user conducts
a search for rich media, they are still inundated with irrelevant
results. Many of those images on the net, Malladi says, are
"dollar images," or images directly tied to transactions. "When
a consumer searches for a product, they have to rely on text
searches, which usually serve up either far too many results,
or none at all." Furthermore, Malladi contended that visual
searches provide a universal search language, one that is independent
of cultural or language variability.
eVision's first product, the eVe SDK, goes beyond text searches, instead relying on visual properties that are automatically extracted from the images and videos based on image processing and pattern recognition techniques. When an image is entered into an eVe database, the image is distilled into representative visual characteristics, based on object, color, texture, and shape. Algorithms are employed that segment an image into distinct object regions and generate scale-independent descriptions of the regions, known as visual signatures. These visual signatures are then organized into an indexing scheme for retrieval. Signatures can be applied to index images, video content, and any digital patterns. This search technique can be applied to the web at large, with pages indexed in a manner similar to that of major search engines.
To establish a media library, image files and video key frames
are analyzed, stored, and indexed into media collections. Given
an example query image, eVe can search either single or multiple
media collections and return a set of sorted images and videos
based on visual similarity. The results are ranked according
to relevance, and can be further refined.
According to eVision, eVe solves one of the major problems
associated with organizing and cataloging visual data by ingesting
and tagging image assets automatically. Traditionally, when
images are cataloged by attaching keywords, the person cataloging
this information has to open up each image one by one and determine
which keywords are appropriate. For thousands of images, this
can be a time consuming and error prone task. To avoid this,
eVe employs Visual MetaTagging, a process that applies text
tags to large groups of visually similar assets. This technique
helps reduce meta-tagging costs and virtually eliminates errors
that grow out of manual tagging. eVe also automatically creates
a representative Visual Vocabulary for large image and video
databases with vast amounts of data.
Going to market, eVision will first target search engines,
providing an extra level of searching capability, and large
enterprises, allowing them to better mine their visual assets.
The company also says that they are in talks with OEMs to integrate
this solution into operating systems and to port it onto chips.
"If you have an eVe ASIC in a digital camera, for instance,
when you take a picture, that image is automatically indexed,
making it searchable at a later time," Peter Giegoldt, VP of
Besides digital cameras, the eVe chip is applicable for scanners,
camcorders, medical imaging equipment, and satellites to generate
automatic searchable signatures as images and videos are captured.
Other markets that eVision will target include the media asset
management market and verticals, such as medical, brand management,
The eVe solution was in beta when we talked to eVision in mid-June,
undergoing testing at North Plains Systems, Ascential Software,
Piocon Technologies, and SGI. General availability for eVe was
scheduled for late June.