Adaptive Multi Scale Document Binarisation Using Vision Mamba
This work addresses readability enhancement in document image analysis, particularly for historical documents, with incremental improvements over existing hybrid models.
The paper tackles document binarization for historical images by proposing a Mamba-based architecture that scales linearly to handle long sequences efficiently, achieving improved performance with concrete gains in metrics like F-measure and PSNR.
Enhancing and preserving the readability of document images, particularly historical ones, is crucial for effective document image analysis. Numerous models have been proposed for this task, including convolutional-based, transformer-based, and hybrid convolutional-transformer architectures. While hybrid models address the limitations of purely convolutional or transformer-based methods, they often suffer from issues like quadratic time complexity. In this work, we propose a Mamba-based architecture for document binarisation, which efficiently handles long sequences by scaling linearly and optimizing memory usage. Additionally, we introduce novel modifications to the skip connections by incorporating Difference of Gaussians (DoG) features, inspired by conventional signal processing techniques. These multiscale high-frequency features enable the model to produce high-quality, detailed outputs.