Scene Text Recognition Models Explainability Using Local Features

Current XAI methods for text recognition are incomplete: Existing explainability approaches for Scene Text Recognition models only provide simple analysis and fail to fully explore different explanation methods, leaving gaps in our understanding of how these systems make predictions.

Global explanations alone are insufficient: Attribution-based methods that highlight important parts of images produce inconsistent and ineffective explanations because they only consider the model's overall context, missing crucial details about individual character recognition decisions.

Local, character-level explanations matter: The paper's proposed STRExp method improves explainability by examining individual character predictions rather than just the image as a whole, providing more granular insight into how text recognition models work.

Explainability requires systematic benchmarking: The research demonstrates the importance of testing explanation methods across multiple datasets and different model architectures to ensure explanations are reliable and generalizable, not just working in isolated cases.

Better explainability supports responsible AI deployment: By making text recognition models more interpretable through local feature analysis, this work helps build trust in AI systems and enables better identification of potential errors or biases in real-world applications like document scanning or accessibility tools.