Irregular Scene Text Detection Based on a Graph Convolutional Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detecting irregular or arbitrary shape text in natural scene images is a significant challenge.
• Existing methods using Convolutional Neural Networks (CNNs) struggle to capture long-range dependencies between text components due to limited receptive fields.

Purpose of the Study:

• To propose a novel and robust method for detecting irregular text in natural scene images.
• To overcome the limitations of CNNs in capturing relations between distant text regions.

Main Methods:

• Utilizing a fully convolutional network (FCN) architecture based on VGG16_BN to detect text components via estimated character center points.
• Employing a graph convolutional network (GCN) to model text line grouping by inferring adjacency relations between text components.

Main Results:

• The proposed method achieves a high text component detection recall rate with fewer non-character components.
• Experiments on ICDAR2013, CTW-1500, and MSRA-TD500 datasets demonstrate the method's effectiveness in handling irregular scene text.

Conclusions:

• The developed method effectively detects irregular text in natural scenes.
• The approach shows promising performance compared to existing algorithms on public benchmark datasets.