Image and Video Analysis

With the availability of digital imaging devices, users world-wide generate and collect vast amounts of image and video data, and digital visual content has become an integral part of our everyday life. Our research targets at the development of novel approaches for exploiting this content efficiently. This includes intelligent search and browsing functionality, as well as techniques for extracting higher-level semantic information directly from the visual content. Our main research topics are:

Image- and Video Retrieval & Browsing

To utilise huge image and video collections, advanced search and browsing functionality is required. We are working on content-based techniques that do not require a manual acquisition of meta-data, with core aspects of our research being:

Similarity-based Search
Detecting Near-duplicate Images and Video Scenes
Robust Sub-image Retrieval
Visual Clustering
Scalability of Image and Video Search

Visual Classification & Concept Detection

The automatic recognition of semantic concepts in image and video data is a core challenge of modern multimedia management, opening applications like intelligent content filters, surveillance, personal video recommendation, or content-based advertisement. We focus on the following research topics in the area:

Auto-Annotation
Visual Learning from Web Sources
Scene Recognition
Pornography Detection

Text in Image Recognition

While optical character recognition (OCR) is well-developed in the document analysis domain, the detection and recognition of textual information in arbitrary image data is still an unsolved research topic. Here, recognition algorithms have to deal with challenges like strongly varying illumination and scale, different forms of distortion, and uncommon fonts. Our work addresses the following topics:

Text Detection
Recognition of Overlay Text
Recognition of Scene Text

Document Image Analysis

In spite of the increasing relevance of electronic documents, the digitalisation and processing of traditional paper-based documents is still a crucial step in many business processes. MADM is developing advanced algorithms for supporting the automated analysis and interpretation of document images with a focus on the following topics:

Layout Analysis

Beyond the digitalisation of pure textual information with Optical Character Recognition (OCR) algorithms, the interpretation of the layout is also crucial for understanding and digitising the content of a document. We are working on basic layout analysis technologies including

Intelligent Document Image Preprocessing Algorithms
Geometric Layout Analysis and Page Segmentation
Logical Layout Analysis and Reading Order Detection

Document Counterfeit Detection

Another focus of our research are novel approaches for verifying the originality of sensitive documents like receipts, certificates, or contracts. Important research topics here are:

Copy Detection
Printing Technique Classification
Printer Identification
Geometric Layout Consistency Verification
Digital Document Fingerprints

Data Mining

Beyond the work in the image and video domain, MADM is also conducting research in other data mining domains.

Pattern Recognition Engineering

One goal of this research is to make pattern recognition algorithms more accessible to real-world software developers which don't have a deep expertise in these algorithms. Research topics here are:

Automated System Evaluation
Automated Model Construction
Meta Learning and Classifier Optimization
Collaborative Development Tools

Anomaly Detection

An important issue in many application domains is the detection of anomalous data instances and data patterns. A popular example is the detection of misuse in financial transaction data or telecommunication systems.

We are developing statistical anomaly detection algorithms with a focus on unsupervised scenarios where neither labeled normal nor labeled anomalous data samples are available for training (semi-)supervised algorithms. Central research topics here are:

Nearest Neighbor based Approaches
Clustering-based Approaches
Scalability
Efficient Application on Real World Data Sets

Network Security

Another goal is the application of statistical pattern recognition algorithms in network security applications:

Detection and Mitigation of Distributed Denial of Service (DDoS) Attacks
Anomaly Detection in Computer Networks

Financial & Business Applications

Moreover, we have expertise in applying pattern recognition and machine learning approaches in financial and business applications. Typical application examples here are credit scoring and fraud detection.