Soft-set MSER Text System

Soft set-based MSER end-to-end system tackles occluded scene text by combining Maximally Stable Extremal Regions (MSER) with soft set theory for robust text detection, recognition, and missing character prediction. Published in Knowledge-Based Systems (2024).

Problem

Scene text in the wild is often partially occluded by objects, shadows, or degradation. Standard detection pipelines miss or mis-recognize such text. Predicting the missing characters requires higher-order reasoning beyond standard sequence models.

Approach

• MSER-based candidate region generation with soft-set-theoretic filtering to reduce false positives
• Recognition module handling partial character sequences
• Soft-set-based prediction head for recovering occluded characters based on contextual soft membership
• Validated on IIIT5K, SVT, and a custom occluded text benchmark

Publication

Das, A., Shivakumara, P., Banerjee, A., Antonacopoulos, A., Pal, U. Soft set-based MSER end-to-end system for occluded scene text detection, recognition and prediction. Knowledge-Based Systems, 2024.