Using GPUs to Improve Image Processing Throughput at FamilySearch

Abstract

Graphics Processing Units (GPUs) have been traditionally used to accelerate computation of computer graphics in applications such as video gaming and high-end 3D rendering. However, recent research has examined using GPUs “in reverse” [1] for computer vision types of image processing. This paper examines leveraging the parallel processing capabilities of GPUs to lower costs and increase the throughput of the millions of original record images being processed by FamilySearch.Digitization of original records is a large focus of archives and family history service providers such as FamilySearch. This digitization enables researchers to more easily access images of records without requiring physical access to archives or microfilm. After digital images of records have been captured, FamilySearch applies several treatments to the raw images to produce both preservation and distribution quality images.Examples of these treatments include decoding from and encoding into different image formats, automatic skew correction, automatic document cropping, image sharpening and image scaling to produce thumbnail images. The intent of these treatments is to enhance the presentation of the records in the images to the end user and to reduce file size for storage.FamilySearch currently processes millions of images annually in this manner through a collection of CPU based servers called the Digital Processing Center (DPC). While the DPC consists of many CPU cores running in parallel across multiple servers, recent GPUs include comparable numbers of less powerful cores in a single card.If servers are constructed with both CPUs and GPUs and code is written to utilize the multitude of cores on the GPUs in a parallel manner, comparable throughput may be achieved in a smaller form factor with less overall cost and decreased processing time per image. The result is increased scalability as FamilySearch continues to increase the number of images processed to make more records available more quickly to more people.