noisy-labels
Accurately segmenting land boundaries from Thai land title deeds is crucial for reliable land management and legal processes, but remains challenging due to low-quality scans, diverse layouts, and complex overlapping elements in documents. Existing methods often struggle with these difficulties, resulting in imprecise delineations that can cause disputes or inefficiencies. To address these issues, we propose CHULA, a novel Custom Heuristic Uncertainty-guided Loss tailored specifically for robust land title deed segmentation. CHULA uniquely combines domain-specific heuristic priors with uncertainty modeling in a unified loss function that effectively guides the model to focus on clearer regions while refining boundaries and suppressing noisy areas. Evaluated on a carefully curated Thai Land Title Deed Dataset, CHULA achieves an impressive 92.4% accuracy, significantly surpassing standard segmentation baselines. Our results highlight the promise of integrating uncertainty and heuristic knowledge to enhance segmentation accuracy in complex, real-world documents. The code is publicly available at https://github.com/kaopanboonyuen/CHULA.
