Research opportunities of the Chair of Information Management

Research result: Automated generation of robust synthetic image datasets and AI models for computer vision applications

Computer-based vision (Computer Vision, CV) offers a wide range of unforeseen, untapped potentials in the production sector. Particularly in the area of quality control, image recognition systems help to identify defects in the production process more quickly and reliably, thereby avoiding or reducing waste. However, manually collecting the image datasets required for this purpose and training the AI models is time-consuming and consequently very costly. To reduce costs and accelerate the adoption of computer vision solutions, HTWD has developed concepts for the automated generation of synthetic training image data. These concepts allow for the generation of photorealistic images of any objects in various environments. They enable the automated and rapid creation of synthetic and labeled image datasets of customer-specific products and processes, which are then used to train neural networks for robust object recognition models. These models are, for example, suitable for extracting meaningful information from images, which can then be used to ensure various optimization measures in the areas of quality control, maintenance, and assembly, thereby easing the burden on employees.

Background: Many tasks in computer vision heavily rely on object recognition, such as object segmentation, image annotation, and object tracking. Convolutional Neural Networks (CNNs) play a key role in object recognition today. However, the performance of CNNs largely depends on the quality and quantity of training datasets, which are often difficult to produce in real-world applications. To ensure the robustness of such models, it is important that training instances are created under various random conditions. These conditions are typically a combination of different factors, including lighting conditions, object location, the presence of multiple objects in the scene, different backgrounds, and the camera angle. Specifically, companies require specialized models for their customer-specific products depending on the application area (e.g., defect detection, anomaly detection, condition monitoring, predictive quality, and so on). Therefore, they often struggle to provide a large number of randomized conditioned instances of their objects. The main reason for this is that the process of capturing randomized and conditioned images of real objects is usually expensive, time-consuming, and practically challenging. Due to the efficiency of synthetic data for training such systems, methods for generating synthetic training data have recently attracted a lot of attention.