Technology that Can Find One Image out of 10,000 in a Second

Searching Large Volumes of Image Data is Laborious

Large volumes of data, including presentation materials and photos, are created and stored every day at offices. Searching for a particular image in this large volume of data often involves a trial-and-error process; users think "Where is the image that I used back then?" or "I think I saw it somewhere around here." Finding the target image is inevitably a time-consuming process. A survey has reported that 10% of work time was spent on searching for data.

"Partial-image search" is looking for images or parts of images that are identical or similar to the image used as a query. The target image is often found with the help of recollection or association. Partial-image search usually requires massive computing resources because comparisons must be made of all parts of all the target images. Such processing in turn imposes a heavy load on a general-purpose server. Search speed may improve if many servers engage in parallel and distributed processing, but this will result in heavy power consumption and high hardware costs.

Focusing on Specific Application Domains to Dramatically Improve Server Performance

Fujitsu Laboratories has developed technology that dramatically accelerates the speed of a partial-image search. With the focus on a specific application domain, this technology, which is based on the "domain-specific server" approach, greatly enhances server performance.

Fujitsu Laboratories analyzed the partial-image search algorithm in detail. After looking at all types of processing required for the search, it relocated the feature extraction and matching processing, both of which cause heavy processing loads, to a general-purpose FPGA.* Then, in line with the characteristics of each type of operation, Fujitsu Laboratories incorporated compact and efficient computing units into the FPGA to achieve high-level parallel processing. Also, scheduling technology that rearranges the processing order according to the situation allows the parallel computing units to work continuously, realizing high-level processing performance.

*: Field-programmable gate array; an IC that can be configured by a purchaser or designer

How partial-image searching is implemented on the FPGA

Quick and Intuitive Retrieval of the Target Image from a Large Volume of Image Data

The processing speed of servers to which the new technology is being applied is at least 50 times faster than general-purpose server using standard technology. The time to search 10,000 images can be reduced from over 1 minute to approximately 1 second. Compared to traditional servers performing at the same level, the power consumption is thirty times lower or more and the size of the device volume is fifty times smaller or more when compared to traditional servers performing at the same level.

The new technology allows users to find the target image quickly and intuitively from the large volume of image data created and stored every day. It is expected to contribute to dramatic operation efficiency improvements. The instantaneous grouping of similar photos allows users to view multiple photos and see an overview of large volumes of stored photos.

Fujitsu Laboratories plans to proceed with development of a system using this high-speed image-search technology with the goal of practical implementation during FY2016. It also plans to expand the technology’s capabilities to allow not only image processing but also the smooth handling of a variety of other types of media data.