Rotated Text Detection via Global Geometry Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current text detection techniques using local text geometry estimations are inadequate for accurately detecting rotated or angled text in images, often failing to fully enclose the text within bounding boxes, leading to suboptimal results.
Innovation Solution
The method involves determining a global text angle based on candidate text angles and using this angle to refine the geometry of bounding boxes, ensuring they accurately enclose rotated or angled text by accounting for the text's rotation, thereby improving detection accuracy and consistency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If local text geometry estimations are used for text detection, then the detection process is simple and fast, but the accuracy of detecting rotated or angled text deteriorates
Solution Approach 1:
The patent segments the text detection task into two parts: local text geometry estimation (for simplicity and speed) and global text geometry estimation (for accuracy on rotated text). By dividing the detection process, the system maintains computational efficiency while improving accuracy for rotated or angled text through the global estimation component.
Solution Approach 2:
The patent merges local text geometry estimations with global text geometry estimations to create a comprehensive detection system. The global estimation results are used to refine and correct the local estimations, combining the speed of local methods with the accuracy of global methods for rotated text detection.
2Reliability
If local text geometry estimations are used, then computational resources are conserved, but bounding boxes fail to fully enclose rotated text
Solution Approach 1:
The patent applies partial action by using global text geometry estimation only for rotated or angled text cases rather than all text. The system first performs local estimation, then selectively applies global estimation when rotation is detected, consuming additional computational resources only when necessary to improve bounding box enclosure accuracy.
3Measurement precision
If traditional text detection methods are used, then processing speed is maintained, but detection consistency for rotated text deteriorates
Solution Approach 1:
The patent performs preliminary local text geometry estimation for all text regions first, which is computationally efficient. Then, based on the preliminary results, it selectively applies global text geometry estimation only to regions showing signs of rotation. This preliminary action maintains processing speed while ensuring consistency for rotated text through targeted refinement.
Data Source
AI summary
Systems, processes and methods for detecting rotated or angled text in an image based on global text geometry estimations are provided. A method includes, at an electronic device with memory and one or more processors, receiving an image including a plurality of pixels (802); determining, based on the image, one or more pixels of the plurality of pixels included in the image that contain text (804); identifying, based on the one or more pixels that contain text, a plurality of components in the image (810); determining a subset of components based on the plurality of components (814); determining, based on the pixels that contain text of the subset of components, one or more candidate text angles (816); determining a global text angle based on the determined one or more candidate text angles (824); and determining a first plurality of bounding boxes based on the global text angle (830).