Reproduce “Primitive Representation Learning for Scene Text Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruijie Yan, Liangrui Peng, Shanyu Xiao, Gang Yao
A new method called primitive representation learning is proposed to extract robust, globally informed visual features for scene text, leading to enhanced recognition accuracy, especially for irregular and multi-oriented texts. This approach also stabilizes attention mechanisms in encoder-decoder models and facilitates efficient parallel decoding.
Key insights
- Stable and efficient feature representations, termed "primitive representations," can be effectively learned by modeling feature map elements as graph nodes and aggregating information globally.
- Combining both pooling aggregators (for intrinsic structural cues) and weighted aggregators (for sample-specific details) yields a more robust and comprehensive feature set for diverse text.
- Integrating purely visual, globally-informed primitive representations as guidance into 2D-attention-based decoders can effectively mitigate the common misalignment problem.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.