Example of GPU Usage at ImageMetry and Verifeyed

A remarkable growth of digital images and videos has been observed in the past decade. The question now is, how much can we trust these images and videos? One of our products is called Verifeyed which is an image and video forgery detection software which can be downloaded from www.verifeyed.com .

Verifeyed is a leading innovator and technology provider for image forensics field. The mission of the product is supporting widespread use of digital photos and videos by increasing their credibility and reliability in the business world through deployment of the latest research results. The technology does not use any watermarks or signatures, which is regarded as a significant technical advantage.

Most of methods we use at Verifeyed are based on complex mathematical and computational methods.For example, a cloning detector, used to identify a part of the image copied and pasted with the typical purpose of hiding an important object or region of the image, operates as follows: (1) Tiling the image with millions of small overlapping blocks; (2) Invariant representation of the overlapping blocks using Fourier-based and Moment-based features [Flusser98]; (3) Kd-tree representation and blocks similarity analysis. Despite the strong detection ability of the cloning detection method, it requires an average several minutes to process typical images which imposes a strong limitation on the usability of the method at sites where thousands of images have to be processed every day.

Another technique detects the altered regions of JPEG images by identifying inconsistencies and change points of the discrete cosine transform coefficients of the JPEG images. The algorithm applies millions of discrete cosine transforms to different sections of the image before transforming the data to a set of one-dimensional statistics and
measuring the distance from the homogeneity conditions (homogeneity conditions are based on JPEG quantization artifacts and image noise levels). To detect the change points, a global approach where all the change points are detected simultaneously is needed; for typical images of 1024×768 this takes around 40 minutes.

Since algorithms in the current Verifeyed system have a strongly parallel nature and thus have great potential for GPU acceleration.  Thus, here, the GPU technique supports major performance improvement of mentioned methods and allow us to use advanced detection techniques and generate results just in a few milliseconds or seconds.

Specifically, we use the GPSME toolkit to rapidly decrease the cost of converting CPU programs into its GPU versions, resulting in noticeably lower execution times for end-users and allowing Verifeyed to bring more cutting-edge image forensics technologies to real-life markets like media, crime investigation, and insurance. This has a significant positive influence to the market position of the company.

Proudly working with: