Student Topics

Symbol Recognition in Images and Videos – Robust Methods for Complex Scenarios

Project/Bachelor thesis
Supervisor: Martin Kampel
Status: open

Motivation
Automated recognition of visual symbols is a central topic across many application areas, including intelligent goods tracking, digital marketing, and the protection of intellectual property and trademark rights. Symbol recognition is also particularly relevant in identifying extremist symbols (e.g., Nazi symbols or terrorist organization logos) on digital platforms for threat prevention and regulation.
A major challenge lies in the high visual similarity between symbols and the growing number of symbol classes. Furthermore, variability in resolution, representation (blurred, printed, digital, hand-drawn), and capture conditions complicates recognition performance.

Objective
The goal of this thesis/project is to evaluate and further develop symbol recognition systems that produce robust results despite high class counts and poor image quality. Based on existing object detection models like Fast R-CNN, YOLO, or DETR, recognition approaches for symbol classes will be analyzed, optimized, and improved through domain adaptation and ensemble methods.
A particular focus is on reducing misclassifications between visually similar symbols — for example, through targeted feature selection, descriptor fusion, or class-specific adaptation strategies.

Tasks

  • Literature review on symbol recognition, domain adaptation, and robust classification

  • Dataset review: selection of suitable symbol datasets (e.g., FlickrLogos, LogoDet3K, symbol sets relevant to Nazism/extremism, if ethically permissible)

  • Application and comparison of current detection models (YOLOv8, Faster R-CNN, etc.)

  • Implementation of domain adaptation strategies (e.g., transfer learning, feature alignment)

  • Testing of ensemble methods (e.g., voting, confidence weighting, descriptor combinations)

  • Evaluation based on precision/recall, class confusion, robustness

  • (Optional) Development of an interactive demo system for symbol recognition in video streams

Required Skills

  • Very good knowledge of Python

  • Experience with deep learning frameworks (e.g., PyTorch or TensorFlow)

  • Basic understanding of computer vision and object detection

  • Interest in real-world applications and ethical considerations

Contact: martin.kampel@tuwien.ac.at