Computer Vision

Overview

Computer vision is an exciting branch of machine learning that aims to teach the machine to process imagery in a way similar to what humans can do. While we are along way from true AI in this regard, computer vision has enjoyed tremendous advances over the past three decades, in which the state of the art has shifted paradigms several times: starting with unsupervised morphological approaches, jumping to hand-crafted supervised learning approaches, and now moving to deep learning networks, automated feature crafting, and semantic object detection, and generative models. The capabilities of the machine here are advancing so quickly that they are now raising new questions in the philosophy community about the true meaning of video, imagery, and authenticity. Many imagery-related tasks and business processes that were thought impossible to automate a decade ago are now possible, and the world is only starting to realize it.

Below are few examples of projects we have worked on that are not confined by NDAs. If you have a vision for an ambitious image processing project that you would like us to take on, please send us a proposal through our contact page, we will be happy to talk with you.

Automated House Detection

QED has trained deep neural networks to perform automated semantic segmentation of households from satellite imagery. This approach is particularly pragmatic for demographic surveillance of rural areas in the developing world, where maps of human settlements are very limited and the logistical costs of conducting annual in-person census surveys in remote regions are often unsustainable. This machine-generated information is of vital importance to the State for planning logistics and optimizing the placement of interventions.  Below are some examples of fully-automated segmentation conducted in Southeast Asia.

original satellite imagery

automated segmentation of households

DARPA Shredder Challenge

The 2011 DARPA Shredder Challenge sought out technologies for the rapid deshredding of shredded documents, inspired by military scenarios in which shredded documents are acquired by boots on the ground, imaged at high-resolution, and sent back to be rapidly deciphered. Dr. Wu assembled a team with former JPL colleagues, dubbed the Herded Rappers — an anagram of “Paper Shredder” — that received an Honorable Mention and ranked 13th out of over 9000 teams. While many teams in the top ten resorted to manual human assembly or Mechanical-Turk-style crowdsourcing solutions (UCSD), our team developed software to address this problem in an automated fashion. The sequence of images directly showcase different steps of our procedure when executed on the first puzzle.

Cell Segmentation

Cell biologists from a bioinformatics lab presented us with the problem of designing an image processing algorithm for the automated segmentation and geometric measurement of cells from slide images — a task formerly relegated to interns and manual labor. These images are heavily corrupted by artifacts and often suffer from suboptimal illumination conditions. The screenshots below showcase our algorithms at work.

Optical Character Recognition

QED has been building solutions to digitally transcribe data from paper-based, handwritten medical registry books found in small-to-medium sized hospitals throughout Sub-Saharan Africa. These books provide a basic accounting of hospital activities, from which one can deduce patient influx rates, ward workloads, and disease prevalence. Such statistics are critical for efficiently allocating resources across hospitals, monitoring emerging threats, and formulating high-level strategies. Below are some pictures of custom-made hardware and software solutions we developed to digitally capture the handwritten data from these books. Our OCR accuracy matches the current state-of-the-art performance in applying deep learning to recognition of real-world handwriting.