Enhancing Scene Text Recognition with Encoder-Decoder Interactive Model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene text recognition is crucial for applications like autonomous driving and smart retail.
• Existing models (CRNN, ViT, PARSeq) face challenges with multi-scale variations, distortions, and complex backgrounds, limiting feature extraction and semantic modeling.
• There is a need for improved scene text recognition methods that can handle these complexities effectively.

Purpose of the Study:

• To propose a novel scene text recognition model, the Encoder-Decoder Interactive Model (EDIM).
• To enhance feature extraction and semantic modeling capabilities in scene text recognition.
• To achieve state-of-the-art performance, particularly on irregular and distorted text recognition tasks.

Main Methods:

• The proposed Encoder-Decoder Interactive Model (EDIM) utilizes an encoder-decoder framework.
• Incorporates a Multi-scale Dilated Fusion Attention (MSFA) module in the encoder for superior multi-scale feature representation.
• Features a Sequential Encoder-Decoder Context Fusion (SeqEDCF) mechanism in the decoder for efficient semantic interaction.

Main Results:

• EDIM was validated on six regular and irregular benchmark datasets and subsets of Union14M-L.
• The model demonstrated superior performance compared to state-of-the-art methods across multiple metrics.
• Significant performance gains were observed, particularly in recognizing irregular and distorted scene text.

Conclusions:

• The Encoder-Decoder Interactive Model (EDIM) effectively addresses limitations in current scene text recognition methods.
• EDIM's novel modules (MSFA and SeqEDCF) enhance feature representation and semantic interaction.
• The proposed model represents a significant advancement in scene text recognition accuracy and robustness.