Text Detection Apparatus Using Cross-Image Weak Text Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing scene text detection methods, such as the sliding window method and connected component analysis, face challenges in achieving high recall rates while minimizing false positives and maintaining real-time performance, and struggle to classify weak texts as strong texts when a significant percentage are detected.
Innovation Solution
A method that detects text candidates in an image by analyzing pixel feature values, classifies weak texts as strong texts by referencing information from another related image, using threshold values and feature conversions, and employs an ER tree and MLBP encoding to ensure high recall and precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the sliding window method is used to detect text candidates, then recall rate is improved, but calculation complexity and false positive rate increase
Solution Approach 1:
The image processing is divided into multiple passes: first extracting text candidates using connected component analysis, then classifying them into strong and weak texts, and finally performing selective refinement. This segmentation allows the system to handle different types of text regions with appropriate methods, improving recall while managing computational complexity.
Solution Approach 2:
Different classification thresholds and refinement strategies are applied to different types of text candidates. Strong texts (those clearly meeting threshold criteria) are handled differently from weak texts (those near the threshold), allowing the system to optimize processing for each category and maintain high recall without uniformly high computational cost.
2Productivity
If connected component analysis is used to extract text candidates, then calculation efficiency is improved, but recall rate decreases due to limited constraint evaluation
Solution Approach 1:
The system performs preliminary classification of text candidates into strong and weak categories before final refinement. This preliminary action allows the system to identify potential text regions that may require further evaluation, preventing premature rejection of valid text candidates and improving recall while maintaining the efficiency of connected component analysis.
Solution Approach 2:
The system uses feedback from the classification stage to guide the refinement process. Weak text candidates are evaluated with additional constraints and information from related images, allowing the system to correct initial misclassifications and improve overall recall rate while maintaining computational efficiency.
3Reliability
If weak texts are classified as strong texts by referencing related images, then recall rate is improved, but system complexity increases
Solution Approach 1:
The system performs preliminary classification of text candidates into strong and weak categories using simple threshold comparisons before applying more complex cross-image analysis. This preliminary action reduces the number of candidates requiring complex processing, improving recall while managing system complexity by applying sophisticated methods only where necessary.
Solution Approach 2:
Cross-image information is applied selectively to weak text candidates rather than all text regions. The system uses local quality assessment to determine which weak texts benefit from additional verification in related images, improving recall for ambiguous cases while avoiding unnecessary complexity for clearly identifiable texts.
Data Source
AI summary
A method for detecting texts included in a specific image is disclosed. The method includes steps of: (a) an apparatus detecting or allowing another device to detect one or more text candidates in the specific image by referring to feature values of pixels in the specific image; (b) the apparatus classifying or allowing another device to classify one or more weak texts in the specific image as strong texts by referring to information on at least one text classified as the strong text in another image related to the specific image if more than a certain percentage of the detected text candidates are classified as the weak texts as a result of comparison between at least one threshold value and at least one feature value of at least one pixel selected in a region where the detected text candidates are included or a value converted from the feature value.