Webrelaunch 2020

Welcome at the Institute AIFB

the Institute of Applied Informatics and Formal Description Methods

Hauptseite/en

Aus Aifbportal
Wechseln zu:Navigation, Suche

Vorlage:Infobox column Welcome at the Institute AIFB Universität Karlsruhe (TH),
the Institute of Applied Informatics and Formal Description Methods.


Professors
Bild-platzhalter.gif
 


News and Events
Events:
2024-04-19: With the growing adoption of Artificial Intelligence (AI) and the successful application of deep learning methods in various domains, AI, and particularly deep learning, increasingly influences people's lives. However, depending on the use case, wrong decisions can be costly and dangerous (e.g., an AI medical diagnosis system misclassifies patients' dis-eases). The emerging topic of Explainable Artificial Intelligence (xAI) offers approaches and algorithms that introduce transparency into black-box models by producing explana-tions of AI Systems' inner workings and decisions. Specifically, in industrial use cases, where complex problems and decision-making processes are widespread, enabling transparent automation and decision support is crucial. However, while research in xAI is trending, applying xAI in industrial use cases is challenging. For many data types (e.g., images or tabular data), xAI methods are well studied. Nevertheless, support for time se-ries, which are ubiquitous in industrial settings, is missing. Further, to use any xAI method in deployment, understanding the explainers' quality, strengths, and weaknesses is of ut-most importance to prevent ambiguous and incorrect explanations. Well-performing xAI methods can help users to understand the reasons behind a deep learners' prediction and enable the recognition of spurious correlations learned by a deep learner or missing information in the collected data. Especially in industrial settings, where only a limited amount of (often) noisy data is available, reverting incorrect model decisions and explana-tions provides the opportunity to include domain knowledge of users. Providing human feedback on the explanation enables a deep learner to infer the missing context and close this gap. To address these application obstacles of xAI in industrial settings, we introduce methods for xAI on time series, the evaluation of xAI, and xAI-based model revisions. at 2:00 pm
2024-07-03: In our partially digitalized world, printed documents are ubiquitous, despite the ongoing efforts to digitalize them. Printed forms play essential roles in various business workflows, such as tracking orders and goods in commercial shipments. To process the information on these forms, data is usually inputted into a computer system through scanning or manual entry. However, both of these methods are time-consuming and inflexible, as there is not always a scanner available nearby. To mitigate this problem, recent research has focused on the development of data extraction systems that can automatically extract data from images of printed forms captured by smartphones. Due to unrestricted environmental conditions and document deformations during the capturing process, the images are often of poor quality, which makes the extraction process difficult. More specifically, the environmental factors include ambient lighting and shadows, whereas document deformations result from the capturing angle and the document's physical condition, such as bends, crumples, folds, and similar factors. In order to improve the extraction process, the images are often enhanced by geometric dewarping and illumination correction. The former aims to remove document deformations, while the latter aims to remove the effects of uneven lighting. In contrast to prior work, we integrate reference template images in the document enhancement process. These reference templates are RGB images of the document in its digital version but without any information on the document's content, thus providing information about the expected layout. The layout information includes the document's structure, such as the position of the texts, structural lines and, logos, as well as the visual appearance of the aforementioned elements. By leveraging the prior knowledge of the document's structure and visual appearance, we improve the geometric dewarping and illumination correction processes. In particular, we propose two approaches for geometric dewarping and one approach for illumination correction, all of which integrate additional reference template information. Thereby, we bring the document images closer to their digital counterparts, which in turn improves the performance of the subsequent data extraction process. at 3:45 pm


News:


Vorlage:Navhighlight