Robust String Text Detection for Industrial OCR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional optical character recognition (OCR) methods often fail due to difficulties in robust line detection and character segmentation, and while sliding window approaches can be more robust, they struggle with false positive detections.
Innovation Solution
A robust string text detection system that avoids segmentation by using a sliding window approach, employing dynamic programming and tree-search algorithms to identify lines of characters and assemble them into likely strings, while filtering out outliers and false positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional line detection and character segmentation is used, then character recognition can be performed, but the system fails when line detection or segmentation fails and segmentation is very difficult to do robustly
Solution Approach 1:
The patent applies segmentation by dividing the image into multiple candidate character detections using a sliding window approach, then segments these candidates into distinct character instances through non-maximum suppression and filtering operations. This allows the system to handle overlapping and varying character sizes robustly.
Solution Approach 2:
Instead of the traditional approach of detecting lines first then segmenting characters, the patent inverts the process by directly detecting character candidates throughout the image using sliding windows, then assembling them into lines and strings. This eliminates the fragile line detection and segmentation steps that previously caused failures.
2Reliability
If sliding window character detection is used to avoid segmentation, then robustness is improved, but false positive detections are generated that must be filtered out
Solution Approach 1:
The patent performs preliminary filtering of candidate character detections by applying multiple constraints before final recognition: spatial consistency checks, aspect ratio filtering, confidence thresholding, and non-maximum suppression. This preliminary action reduces false positives while preserving true character detections.
Solution Approach 2:
The system uses feedback mechanisms where detected character candidates are evaluated against expected character properties (size, spacing, orientation, confidence scores), and detections that不符合 these feedback criteria are filtered out. The system iteratively refines detections based on this feedback to eliminate false positives.
Data Source
AI summary
A robust string text detection and recognition system is provided to improve the reading performance of sliding-window based OCR solutions (such as histogram of oriented gradient (HOG) OCR and convolutional neural network (CNN) OCR). A search algorithm is created that is robust enough to detect outliers and false detections. A general Text Search algorithm structure is created allows the specification of different constraints/assumptions to guide the search in multiple text lines detection and recognition.