Scene Text Recognition Using Segmentation and Relationship Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current scene text recognition (STR) methods face significant challenges due to complex image backgrounds and diverse fonts, leading to difficulties in accurately recognizing text in natural scenes.
Innovation Solution
A method and apparatus that utilize text instance segmentation to generate feature maps, construct a relationship graph with nodes representing pixels and edges based on spatial semantic features, and process this graph using a pre-trained graph convolutional network to extract two-dimensional spatial features for improved text recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional text recognition methods are used for Scene Text Recognition, then the recognition process can be simplified, but the recognition accuracy deteriorates due to complex image backgrounds, diverse fonts, and random text arrangements
Solution Approach 1:
The patent applies segmentation by dividing the text recognition task into distinct stages: text detection to locate text regions, text instance segmentation to separate individual text instances, and recognition to identify the text content. This multi-stage segmentation approach handles complex scene text by breaking down the challenging task into manageable steps, improving accuracy without overwhelming system complexity
Solution Approach 2:
The patent introduces a new dimension by constructing a relationship graph that captures spatial semantic relationships between text pixels. Instead of processing text as simple 2D images or 1D sequences, the method creates a graph structure that adds relational dimensionality, enabling the model to understand spatial arrangements and semantic connections between different parts of text in complex scenes
2Measurement precision
If text instance segmentation is performed to preserve two-dimensional spatial information, then text recognition accuracy improves, but computational complexity increases
Solution Approach 1:
The patent extracts only the essential spatial semantic information needed for recognition by constructing a relationship graph that selectively captures relevant pixel relationships. Instead of processing all possible spatial information, the method extracts and processes only the critical spatial semantic features, reducing computational energy while preserving the two-dimensional spatial structure necessary for accurate recognition
Data Source
AI summary
A method for recognizing a text, and an apparatus. A specific embodiment of the method comprises: obtaining feature maps, the feature maps being obtained by means of performing text instance segmentation on an image having a text to be recognized; constructing a relationship graph according to the feature maps, wherein in the relationship graph, each node represents a pixel in a feature map, and each edge represents that a similarity measure of spatial semantic features of two connected nodes is greater than a target threshold, and a spatial semantic feature of a node comprises a type feature and a position feature of a pixel represented by the node; utilizing a pre-trained graph convolutional network to perform processing on the relationship graph, and obtaining a first text feature corresponding to the image; and generating a text recognition result for the image according to the first text feature.