COMET-Project deepTrust (2019-2022)

Deep Models in Trust Critical Image Analysis Systems

Motivation

This project focuses on developing and advancing analysis methods towards more trustworthiness and transparency in deep learning driven image analysis systems. Explanatory factor analysis and attention mechanism based approaches are studied to tackle the problem of distinguishing nuisance from task-relevant data and features, and to gain a better model understanding. A proof of concept will be addressed by image analysis use cases in trust-critical application domains of medical imaging and security in terms image-based identity verification.

In particular, this project is motivated by the following industrial use cases:

  • diagnostic decision support for physicians in ultrasound imaging
  • detection of fraudulent manipulations of items of identification (ID cards, passports): Validation of a person’s identity is an important task in many online applications, e.g., shops, services

Expected Results

The goal is to devise best-practice tools and workflows for trust analysis of deep learning based image analysis systems as a basis for model and data debugging and novel approaches for testing. 

Other expected results are:

  • application-specific benchmark datasets for trust analysis from ultrasonic medical imaging and identity verification based on ID documents
  • Python software prototype for software framework for benchmarking trust analysis methods
  • Python software prototype of universal model approach and its proof of concept
  • Python software prototype of multimodal learning approach and its proof of concept
  • scientific analysis of outlined approaches regarding universal model and multi-learning approaches
  • software dashboard for trust evaluation in terms of data and model quality and transparency
  • experimental analysis and review of current software packages in this field
  • peer-reviewed scientific publications, e.g., IEEE transactions on industrial informatics

Funding Partner

This project is subsidized in the frame of COMET – Competence Centers for Excellent Technologies by BMK, BMDW, State of Upper Austria and its scientific partners. The COMET program is handled by FFG.

Contact

Lukas Fischer

Fischer Lukas

Research Manager Data Science
Phone: +43 50 343 828

Thomas Hoch

Hoch Thomas

Researcher Knowledge-Based Vision Systems
Phone: +43 50 343 831

back