Three dimensionalization techniques for epipolar views, and design process reconstruction.

Companies which straddled the digital revolution, often have a large quantity of data still stored in physical forms. Digital data is more readily accessible for analysis and processing, less liable to deterioration, and cheaper to store: thus, transferring relevant information from hard to soft copies of documents is becoming a pressing issue. Industries that designed products before the digital revolution, and which have this product still on the market, have an interest in the migration of this data to reduce time burden and costs. A feasibility study, funded by AIRBUS, is aiming to produce code which can extrapolate data in digital form from a lot of images. Approaching the problem with a more generalised method could enable more than a hundred thousand images to be processed without requiring human interaction. The greater the automation of this process, the lower the cost will be - as long as the reliable information is retained. Technologies evolve, and quickly editable robust parameterization of design will enable industries to cope with rapid changes of design intent. Different methods have been developed to extract entities and given their non-commutative nature the results changes depending on the order in which such operations are carried out. Given the large number of drawing that will be iterated the process must be coupled with a pattern-matching deep learning algorithm which would aim to assess the order of function yielding the best result, after having assessed the prevalence of individual entities within a drawing, regarding the number of entities and quantity of pixels. The variance of noise in such broad population of images is quite large. Therefore, robust and adaptive filters will be necessary. Using non-fully converged geodesic active contours to remove noise will be attempted, this will also be used to create masks for inactive part of the drawings which do not require processing. A similar approach can be applied to other industries with information stored in images ranging from medical imaging to a large number of cadastral maps. Often this kind of data needs to be accessed from a large number of people and different teams. Digitalization would improve collaborations and reduce time and costs necessary for accessing such data, resulting in greater productivity. Additionally, access to larger data sources would produce more informed solutions. After this first stage of digitisation, we aim to explore methods for 3D reconstruction of manufacturing designs. The difference from ordinary 3D reconstruction is that most drawings are designed so that two views carry enough information for an object to be manufactured, hence be reproduced in 3D. The research will, therefore, aim to use extracted 2D features to test a robust 3D reconstruction algorithm based on 2 or 3 views characterised by an orthogonal epipolar constraint. Possibilities method explored will be:
- Construction of hierarchical trees based on the order of operations carried to reconstruct a 3D model, then assessed by consistency checks with original 2D information (computationally expensive)
- Layer-by-layer reconstruction, using cutting plane rendering technique which uses planes on one view and progressively associated to layers on another view moving from front to back. Belonging to an entity to the same plane will require consistency between different possibilities closed loops on one view with another one.

The reconstruction will need to consider design intent, as in development the model might have to undergo various changes. My research, supervised by Dr Patterson, would focus on algorithms for image processing and image understanding, aiming to produce working software which can digitalise a large quantity of data with accuracy. In the end, it will seek to reconstruct 3D technical drawing file compatible with most used software (e.g., Autodesk).