Transformer Text Detection for User-Specific Output Shapes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR methods fail to detect text in images due to limitations in selecting appropriate formats and detecting text in images, which are not user-specific or performance-dependent.
Innovation Solution
A method and device for detecting text in images using a transformer model that includes an encoder and decoder.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional OCR methods are used for text detection, then text recognition can be performed, but the detection format cannot be customized according to user requirements or image characteristics
Solution Approach 1:
The system dynamically selects detection formats (bounding box, polygon, center point) based on user requirements and image characteristics, transforming a static OCR system into a dynamic one that adapts to different detection needs without requiring multiple fixed systems
Solution Approach 2:
The text detection model is designed to perform multiple detection formats using the same underlying architecture, allowing a single system to handle bounding box detection, polygon detection, and center point detection, thereby achieving multi-functionality without proportionally increasing complexity
2Loss of information
If a fixed text detection model is used, then the system structure remains simple, but the output information amount is limited and cannot detect text at user-desired locations or in desired shapes
Solution Approach 1:
The system extends detection beyond the traditional bounding box dimension by introducing polygon vertices and center point coordinates as additional dimensional representations, enabling more precise text location and shape detection while maintaining model efficiency through selective dimension usage
3Ease of operation
If conventional text detection is performed without user-specific conditions, then the process is simple, but the detection results cannot be visualized or output according to user-preferred locations, areas, or shapes
Solution Approach 1:
The system pre-processes the input image to generate feature maps and detects all potential text regions in advance, then selectively outputs results according to user conditions, avoiding repeated processing and reducing overall time loss while maintaining user control flexibility
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to a method for detecting text in an image. The method for detecting text in an image comprises the steps of: receiving an image including text; receiving a command including a text detection condition; and inputting the image and the command into a text detection model so as to generate a sequence indicating a detection result of a text instance included in the image according to the text detection condition.