DeepMine-Mamba: Mitigating Information Dilution in Mamba-Based State Space Models for Document Image Binarization
Researchers have introduced DeepMine-Mamba, a new framework for document image binarization that utilizes Mamba-based state space models. The proposed method addresses information dilution issues inherent in direct state-space propagation, particularly for faint or fragmented text strokes. A novel Anti-Dilution Gate is incorporated to selectively restore local responses sensitive to stroke details while mitigating background enhancement, leading to competitive performance on DIBCO/H-DIBCO benchmarks. AI
IMPACT Introduces a novel approach to document image binarization using Mamba models, potentially improving performance on degraded documents.