Publication
From Laboratory State into Production: Automating Hyperspectral Image Classification through MLOps
Lena Herrmann; Alina Griesel; Carmen Moßner; Jan-Philipp Schwarze; Martin Atzmueller
In: Proceedings 2025 IEEE 37th International Conference on Tools with Artificial Intelligence ICTAI 2025. International Conference on Tools with Artificial Intelligence (AAAI SSS-2025), November 3-5, Athen, Greece, Pages 512-519, IEEE Xplore, 12/2025.
Abstract
When developing AI-based applications, it is essential to create an environment that follows modern standards of software engineering. This is equally important for both research-based and industry-based AI development, which also includes advancing from a laboratory state to production. The development process of AI systems still faces challenges like difficulties in deploying models, tracking and comparing experiments or collaporating on projects. Furthermore, repetitive steps lack automation which leads to opaque, inefficient and errorprone workflows. Through the emergence of MLOps, principles and guidelines were introduced which support the design and implementation of efficient, traceable and transparent development processes. In this regard, we implemented a Kubernetes-based MLOps infrastructure that provides the functionality to create and run ML applications that reach a TRL 6. We are demonstrating the usability of our MLOps infrastructure through a practical use case for food quality classification. The use case includes the implementation of an automated machine learning pipeline and shows its reusability and robustness through the usage of four different data sets based on hyperspectral images. The ML pipeline consists of pre-processing, dimensionality reduction and CNN model training steps. The dimensionality reduction includes K-Means and Segmented Autoencoders, while CNN model training involves hyperparameter tuning. With this design, it is possible to leverage the best-fitting variant for a given data set. Furthermore, through the level of automation, the entire process can be triggered automatically when new data is collected. Hence, the entire pipeline is usable within a continuous training process and creates an adaptable hyperspectral image classifier. We argue that our established MLOps infrastructure provides an efficient sharing of hardware and software resources and moves the overall development process towards modern standards of software engineering.
Projects
- Movi-Q - Mobile visual quality recognition through artificial intelligence for the food industry
- RLA - AI Real-World Laboratory for Agriculture