Unsupervised Data Extraction from Graphs and Data Plots

Quintessential technical information and IP is often presented in the form of graphs and dataplots (images). These images are usually embedded in one of many possible document types and formats including presentation slides or whitepapers. Especially in large corporations or research institutions the number of documents containing this type of image data is vast. Extracting the underlying data correlations, parametric dependencies, and trends contained in these images in a fully automatic and large-scale fashion would open up new avenues to mine and understand crucial pre-existing domain knowledge. Many day-today engineering activities such as equipment troubleshooting, DOE planning and excursion containment would benefit and be enhanced by enabling experts to discover multi-echelon causal relationships between control parameters, process variables and their effects. The main goal of this project is to develop deep learning models and algorithms that extract data in a structural form from graph and data plots and to incorporate these models into existing knowledge-management systems that focus on text analysis only.