OCR Character Selection Using CTC Coordinate Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optical character recognition (OCR) technologies for smart lenses on mobile devices lack character-level selection accuracy and efficiency, often failing to accurately segment words from punctuation or handle complex backgrounds, and require inefficient pixel-by-pixel traversal for boundary detection.
Innovation Solution
A method utilizing a Connectionist Temporal Classification (CTC) sequence to calculate character coordinates, fine-tune boundaries, and map them to the original picture for precise character selection, supported by an OCR detection and recognition model, enabling character-level selection with improved accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If manual character selection is performed by tapping and dragging on recognized text, then character selection functionality is provided, but character positioning accuracy deteriorates and selection efficiency is reduced
Solution Approach 1:
The patent introduces character coordinate information as an intermediary between the OCR recognition result and the display position. By calculating and mapping character coordinates from the CTC sequence to the original picture coordinate system, the system achieves accurate character positioning without requiring complex pixel-by-pixel traversal, thus resolving the contradiction between providing selection functionality and maintaining positioning accuracy
Solution Approach 2:
The patent performs preliminary calculation of character coordinates during the OCR recognition process itself. By pre-calculating the mapping between CTC sequence indices and picture coordinates before user interaction, the system prepares the positioning data in advance, enabling accurate and efficient character selection when the user performs tapping and dragging operations
2Measurement precision
If pixel-by-pixel traversal is used for boundary detection, then character boundary accuracy is improved, but processing time increases
Solution Approach 1:
The patent replaces the mechanical pixel-by-pixel traversal approach with a mathematical coordinate mapping method. Instead of iteratively checking each pixel to detect boundaries, the system uses the pre-calculated character coordinates from the CTC sequence and applies transformation formulas to directly obtain the boundary positions in the original picture, significantly improving processing speed while maintaining accuracy
Solution Approach 2:
The patent changes the parameter representation from pixel-level data to coordinate-level data. By working with character coordinates derived from the CTC sequence rather than individual pixel values, the system transforms the problem from a computationally intensive pixel traversal task to a efficient coordinate transformation task, achieving both speed and accuracy
3Measurement precision
If character-level selection is implemented, then selection precision is improved, but device complexity increases
Solution Approach 1:
The patent makes the existing CTC sequence processing infrastructure serve multiple functions. The same CTC sequence that provides character recognition also provides character positioning information through coordinate mapping. This multi-functional use of the CTC sequence enables character-level selection precision without requiring separate complex systems for recognition and positioning
Solution Approach 2:
The patent merges the character recognition process and the character positioning process into a unified workflow. By calculating character coordinates during the same CTC decoding process that generates recognition results, the system combines two functions that could have been separate, reducing overall system complexity while achieving character-level selection precision
Data Source
AI summary
Embodiments of this application are applicable to the field of artificial intelligence technologies, and provide a method and an apparatus for character selection based on character recognition, and a terminal device. The method includes: obtaining a connectionist temporal classification sequence corresponding to text content in an original picture; calculating character coordinates of each character in the connectionist temporal classification sequence; mapping the character coordinates of each character to the original picture, to obtain target coordinates of each character in the original picture; and generating a character selection control in the original picture based on the target coordinates. The character selection control is used to indicate a user to select a character in the original picture. By using the foregoing method, when the user manually selects the character, precision of positioning the character can be improved, and efficiency and accuracy of manually selecting the character can be improved.