Deep-Learning technologies is one of the most impressive technology breakthrough in computer science during the last decade.
While the underlying theory has been proposed during the last quarter of the twentieth century, availability of “low-cost” massively parallel computers, namely the GPUs (Graphical Processing Units) has transformed this “theory” into an impressive aperture towards previously unreachable objectives.
From its creation in 1986 DELTATEC business has been closely related to the image, in the industrial, video broadcast, entertainment, medical and space markets. Deep-Learning has demonstrated its stunning capabilities in these domains, enabling creation of machines capable to recognise their environment.
Therefore, DELTATEC has naturally integrated this technology into its solution workflow for pattern and object recognition or image geometry analysis.
For instance, DELTATEC has integrated these technologies in the latest releases of DELTACAST sports systems to support automated operation.
Range of applications inside DELTATEC is constantly growing. New developments include image distortion analysis or mechanical profile monitoring. Deep Learning is a driver in development of the DELTATEC vision business-unit activities.
Deep-Learning techniques are now invading the embedded device world through FPGA designed neural networks and specific devices integrated in SOC processors. DELTATEC ability to manage these technologies in an integrated way enforce its position as a premium player in this area.
Tools & Methods
DELTATEC has developed a comprehensive tool chain for the development of Deep Learning neural networks based on the CAFFE framework by Berkeley Artificial Intelligence Research (BAIR).
This framework has been demonstrated as one of the most stable and computationally efficient in applications.
Particularly, DELTATEC tool chain integrates:
- A set of GPU powered workstation with the NVidia DIGITS tooling for “code free” development of image processing networks
- A DELTATEC adaptation of the original framework for execution using the Windows Operating system
- An integration library that provides a very simple integration path inside applications, separating completely the design of neural networks of the programming of the applications.
Of course, DELTATEC brings attention to the new frameworks that exist in the research community, being ready to integrate them when they reach the required maturity and efficiency levels.
Another aspect of Deep Learning is data collection. Data collection is a fundamental key for Deep Learning because these systems are trained from examples. Large data sets are not always easily available in industry, particularly when designing new products. DELTATEC is working actively with image synthesis to emulate these data sets, providing an efficient way to circumvent this limitation.
Image synthesis can be performed using standard 3D and 2D image generation software. It can also be performed using specific emulation or synthesis tools. For instance, DELTATEC has designed a soccer image generator, based on the DELTACAST Virtual View software, or a lens distortion emulator.
DELTATEC also pay a particular attention to true live environment, training neural networks to make them effective in adverse conditions and testing them with especially realistic test sets.