Algorithmic evaluation demonstrates superior detection and transcription performance across irregular scene text benchmarks, indicating character-level guidance improves visual robustness.
Key Points
To enhance the detection and recognition of arbitrarily shaped, low-contrast, or cluttered scene text by introducing explicit character-level visual and semantic guidance.
Designed CADet for text detection, combining a text-enhancement network (TENet), a character information adaptive guidance module (CIA), and a position/classification compensation module (COMP).
Constructed SIETR for text recognition, combining a character local image embedding module (CLIE) with permutation language modeling (PLM) for autoregressive decoding.
Evaluated detection performance on ArT, Total-Text, and CTW1500, and recognition accuracy across multiple irregular text benchmarks.
CADet achieved F-measures of 79.5% on ArT, 89.4% on Total-Text, and 89.2% on CTW1500, outperforming representative Transformer-based detectors.
SIETR reached a 95.6% sample-size-weighted average accuracy with 23.8 M parameters and 3.2 G FLOPs, exceeding PARSeq performance on irregular text benchmarks with fewer FLOPs.