The paper introduces a search-based weight repair strategy for regression-based DNNs, enabling targeted weight adjustments for performance improvement of semantic segmentation — overcoming the limitations of traditional classification-based methods and advancing robust model correction in safety-critical AI systems.

Filter-based Repair of Semantic Segmentation in Safety-Critical Systems
Sebastian Schneider¹, Tomas Sujovolsky², Paolo Arcaini³, Fuyuki Ishikawa³, Truong Vinh Truong Duy⁴
1 Technical University of Munich
2 University of Buenos Aires
3 National Institute of Informatics
4 Aisin Corporation

Related link:
URL: https://www.computer.org/csdl/proceedings-article/saner/2025/351000a349/26TIqPoKuGI
DOI: 10.1109/SANER64311.2025.00040