Curved Text Recognition via Boundary Key Points and TPS
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Optical Character Recognition (OCR) technologies face low recognition rates for curved texts due to the use of quadrilaterals as detection results, which often fail to encompass curved text, leading to inefficient recognition.
Innovation Solution
A method involving a preset feature extraction network and segmentation network to extract feature maps, determine boundary key points, and convert curved texts into a target arrangement sequence using Thin Plate Spline (TPS) for improved recognition processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If quadrilateral text boxes are used for text region detection, then the detection process is simple and fast, but curved texts cannot be fully contained leading to low recognition rate
Solution Approach 1:
The text region detection is segmented into multiple stages: first detecting rough text regions using quadrilateral boxes, then further segmenting and refining these regions to accurately contain curved texts. This multi-stage segmentation approach maintains the simplicity of initial detection while improving the accuracy of curved text containment through subsequent refinement steps.
Solution Approach 2:
The patent transitions from two-dimensional quadrilateral bounding boxes to three-dimensional polygonal representations that can accurately envelop curved text regions. By adding dimensional flexibility (allowing for non-convex and irregular polygon shapes), the system can properly contain curved texts while maintaining computational efficiency through automated polygon generation algorithms.
2Loss of information
If quadrilateral bounding boxes are used to detect text regions, then the detection algorithm is simple, but curved texts are not fully contained resulting in information loss
Solution Approach 1:
The patent applies curvature-based polygon generation that adapts to the curved nature of text regions. Instead of forcing straight-edged quadrilaterals, the algorithm generates polygons with curved edges that follow the contours of curved texts, ensuring complete containment while preserving all text information. This curvature adaptation is achieved through mathematical transformations that map detected text regions into appropriate polygonal forms.
3Reliability
If character-level labeling is performed for curved text recognition, then recognition accuracy may improve, but the processing complexity and time increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-processing curved texts into standardized linear arrangements before recognition. Through automated polygon detection and text straightening algorithms, the system transforms various curved text formats into uniform linear sequences that can be directly fed into recognition models. This preliminary transformation eliminates the need for time-consuming character-level segmentation and labeling during the recognition phase, significantly reducing processing time while maintaining accuracy.
Solution Approach 2:
The patent introduces an intermediary processing stage that acts as a mediator between curved text detection and recognition. This intermediary module includes algorithms for text region segmentation, boundary detection, and text straightening that convert curved texts into standardized formats. By inserting this intermediary transformation layer, the system avoids the complexity of character-level labeling while ensuring recognition accuracy through proper text normalization.
Data Source
AI summary
Methods and an apparatuses for recognizing a text, recognition devices and storage media are provided, which belong to the field of text detections. A method includes: extracting, by the recognition device, a feature map of a to-be-recognized image, then determining segmentation information of a text region of the to-be-recognized image based on a preset segmentation network and the feature map, and then determining boundary key points in the text region based on the segmentation information, and then converting a text in the text region into a text with a target arrangement sequence based on the boundary key points and then inputting the text obtained by conversion into a preset recognition model for recognition processing.