CDistNet: Perceiving Multi-Domain Character Distance for Robust Text Recognition

The Transformer-based encoder-decoder framework is becoming popular in scene text recognition, largely because it naturally integrates recognition clues from both visual and semantic domains. However, recent studies show that the two kinds of clues are not always well registered and therefore, feature and character might be misaligned in the difficult text (e.g., with rare shapes). As a result, constraints such as character position are introduced to alleviate this problem. Despite certain success, a content-free positional embedding hardly associates stably with meaningful local image regions. In this paper, we propose a novel module called Multi-Domain Character Distance Perception (MDCDP) to establish a visual and semantic related positional encoding. MDCDP uses positional embedding to query both visual and semantic features following the attention mechanism. The two kinds of constrained features are then fused to produce a reinforced feature, generating a content-aware embedding that well perceives spacing variations and semantic affinities among characters, i.e., multi-domain character distance. We develop a novel network named CDistNet that stacks multiple MDCDPs to guide a gradually precise distance modeling. Thus, the feature-character alignment is well built even various recognition difficulties presented. We create two series of augmented datasets with increasing recognition difficulties and apply CDistNet to both them and six public benchmarks. The experiments demonstrate that CDistNet outperforms recent popular methods by large margins in challenging recognition scenarios. It also achieves state-of-the-art accuracy on standard benchmarks. In addition, the visualization shows that CDistNet achieves proper information utilization in both visual and semantic domains. Our code is given in https://github.com/simplify23/CDistNet.