CC BY 4.0Kronenwett, FelixFelixKronenwettLehmann, RomanRomanLehmannZheng, HaoxiangHaoxiangZhengMaier, GeorgGeorgMaierLängle, ThomasThomasLängleKarl, WolfgangWolfgangKarl2025-04-142025-04-142025https://doi.org/10.24406/publica-4530https://publica.fraunhofer.de/handle/publica/48649410.24406/publica-4530In recent years, the demand for efficient and accurate sorting solutions across various industries has surged due to the need for enhanced material recovery and sustainability. Sensor-based sorting systems have emerged as pivotal technologies. They employ visual inspection to ensure the precise classification and sorting of bulk materials. Several challenges hinder the potential of deep learning models in industrial systems for image data analysis in complex sorting tasks. Due to different hardware, traditional static deep learning models often fail to handle the dynamic requirements of varying material throughput and execution times, leading to inefficient sorting accuracy. This paper investigates the integration of adaptive architectures for semantic segmentation. These architectures dynamically adjust their computation pathways based on input complexity, optimizing performance and resource utilization. Implementing an architecture with early exit mechanisms improved accuracy, enabling sorting decisions regardless of hardware limitations. Experimental validation using real-world data from sorting plants demonstrates adaptive models’ practical applicability and benefits for sensor-based sorting systems.enAdaptive architecturessemantic segmentationsensor-based sortingearly-exit mechanismsconference paper