AI/Machine Learning Expertise
SSSC has been active in implementing Deep Learning and Machine Learning for a variety of specific problems of interest to our customers and then deploying the trained networks to either on-premises applications or to the Cloud in a secure environment.
It has been our experience that subject matter expertise has been critical in the successful implementation of Deep Learning algorithms. We use both subject matter and Deep Learning experts in the development of our AI/ML solutions. Although each project is unique, there is a common process to all the projects:
- Problem Definition: A meeting takes place where the problem, timeline and availability of data is discussed
- Data Formatting: Available datasets are formatted and or created for the project
- Preprocessing: The dataset is preprocessed to present the data in the most advantageous way for characterization and/or operation.
- Network Architecture: The processing chain is designed. This constitutes both workflow and network design
- Network Training: The network is trained to increase the accuracy of the results
- Test of Network: The network is tested and improved through iteration
- Deploy Network in relevant environment: The network is deployed at the customer site
SSSC typically uses the Python Programing language for most network implementation. We are well versed in common libraries such as PyTorch and Keras. However, where required, applications can be written in C.
Some examples are shown below:
Machine Learning Applied to Measurement Assessment
The goal of program is the development of a machine learning tool that automatically detects errors and contamination in radar cross section (RCS) measurements at an outdoor test range. This tool exploits recent advances in machine learning by using a deep learning neural network to extract sophisticated sets of features, which is adapted to reveal subtle patterns in the data. These patterns are subsequently used to classify between clean sets of RCS measurements versus those containing radar interference, or other sources of error.
The goal of program is the development of a machine learning tool that automatically detects errors and contamination in radar cross section (RCS) measurements at an outdoor test range. This tool exploits recent advances in machine learning by using a deep learning neural network to extract sophisticated sets of features, which is adapted to reveal subtle patterns in the data. These patterns are subsequently used to classify between clean sets of RCS measurements versus those containing radar interference, or other sources of error.
Automated Synthetic Shortwave Infrared Image Creation From Near Wave Infrared and Electro Optical Imagery
This effort demonstrated the feasibility of an innovative and automated workflow that ingested electro-optical visible-light imagery and Near Infrared (NIR) satellite imagery and produced synthetic Shortwave Infrared (SWIR) images in an Open Geospatial Consortium (OGC) GEOTIFF compliant data format standard for use with Infrared (IR) scene projectors. This effort explored a wide range of methods and techniques for generating synthetic SWIR images. Each was trained using satellite imagery where SWIR image ground truth is available with visible and NIR reference imagery. Experiments were conducted to assess the fidelity of the synthetic imagery, the runtime of the workflow, the adaptability of the algorithm to images of various sizes and formats, and the compatibility with the existing scene generation workflow. This effort used open source tools in all aspects of the workflow. It also employed open, international standard data formats for representing the imagery at all points throughout the processing chain. This will ensure that the automated workflow supported integration with other tools in the future.
Imaging Techniques for Passive Atmospheric Turbulence Compensation
In this work, Solid State Scientific Corporation (SSSC) developed a unique technique for imaging through atmospheric turbulence. The idea combined the use of specially designed hardware and associated algorithms to yield improved performance over other systems. The solution improved upon one of the traditional techniques for imaging through atmospheric turbulence, the so-called, Lucky Look technique.
The approach enhanced the field of medium and high-altitude intelligence, surveillance, and reconnaissance. The approach also improved imaging performance in highly turbulent environments. Key applications for this technology include long range imaging and intelligence, surveillance, and reconnaissance. Commercial applications include imaging from airborne platforms for law enforcement as well as high-end videography and astronomy.
Interesting in learning more / working with SSSC on this effort ?
Email us at: deep-learning@solidstatescientific.com
or Complete the contact form: Contact Us
Email us at: deep-learning@solidstatescientific.com
or Complete the contact form: Contact Us