Faculty of Informatics / Mathematics

[Translate to English:] Symbolbild Serverraum

Automated generation of robust synthetic image datasets and AI models for computer vision applications (ASYN4CVA)

Object recognition is a subtask for computer vision (CV) in which instances of visual objects are assigned to a specific class in digital images. Numerous CV tasks are highly dependent on object recognition, e.g. object segmentation, image labeling and object tracking. Convolutional neural networks (CNNs) currently play a key role in object recognition. The performance of CNNs largely depends on the quality and quantity of training data sets, which are often difficult to produce in real applications. To ensure the robustness of such models, it is important that the training instances are created under different random conditions. These conditions are a combination of various factors, including lighting conditions, object location, the presence of multiple objects in the scene, different backgrounds and the angle of the camera. In particular, depending on the application area (e.g. defect detection, anomaly detection, condition monitoring, predictive quality, etc.), companies need specific models for their customized products. However, they have difficulties in providing a large number of randomized conditioned instances of their objects. The main reason for this is that the process of acquiring randomized and conditioned images of real objects is costly, time-consuming and difficult in practice. Due to the efficiency of synthetic data for training such systems, methods for generating synthetic test data have recently attracted much attention.

The aim of the project "Automated generation of robust synthetic image data sets and AI models for computer vision applications" (ASYN4CVA) is to develop a consistent process for the automated generation of synthetic data and the development of a robust object recognition model for customer-specific products. This process should make it possible to automatically generate robust object recognition models for various industrial applications. The models trained in this way are suitable, for example, for implementing object recognition for user guidance for maintenance and servicing measures on industrial plants or for creating object recognition models for quality assurance in industrial processes to detect missing, incorrectly assembled or additionally assembled parts. Other areas of application for these models include the image-guided control of industrial robots for pick & place or assembly tasks. The process is used to generate synthetic image data based on digital models or real objects in an automated workflow and train CNN models with it.


Project duration

December 2022 to November 2025