Overlap-Aware Optical Character Recognition Using Deep Learning Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing optical character recognition solutions are inefficient and ineffective in detecting text when characters have overlaps, requiring substantial training data and processing resources and often failing to accurately identify separate words in overlapping text.
Innovation Solution
The use of deep character overlap detection machine learning models that generate character maps, overlap maps, and affinity maps, allowing for image-wise arithmetic operations to perform overlap-aware optical character recognition, reducing the need for extensive training data and processing resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing optical character recognition solutions are used to detect text with overlaps, then they require substantial training data and processing resources, but they are still ineffective and fail to accurately identify separate words
Solution Approach 1:
The patent segments the text detection task into three distinct maps: character map (identifying individual characters), overlap map (detecting overlapping regions), and affinity map (determining word boundaries). This segmentation allows the system to process overlapping text more effectively by handling character identification and word boundary determination as separate tasks, improving accuracy without requiring extensive training data for each specific scenario
Solution Approach 2:
The affinity map serves as an intermediary that bridges character-level detection and word-level recognition. By introducing this intermediate representation that captures spatial relationships and affinity between characters, the system can accurately determine word boundaries even in overlapping text without needing substantial training data for word boundary classification
2Measurement precision
If existing optical character recognition solutions are used to detect text with overlaps, then they require substantial processing resources, but they fail to accurately identify separate words
Solution Approach 1:
By dividing the recognition task into three specialized maps (character, overlap, affinity), each map can be processed independently with optimized algorithms. This segmentation reduces computational complexity compared to using a single complex model that attempts to handle all aspects of overlapping text recognition simultaneously, thereby reducing processing resources while improving accuracy
Solution Approach 2:
Each map focuses on a specific aspect of text recognition with specialized processing: character map for individual character detection, overlap map for detecting overlapping regions, and affinity map for word boundary determination. This local specialization allows each component to be optimized for its specific function, improving overall accuracy while minimizing processing resources for each task
3Reliability
If existing optical character recognition solutions are used on overlapping text, then they may produce results, but they are ineffective in accurately identifying separate words
Solution Approach 1:
The patent improves reliability by segmenting the recognition process into three distinct maps that can be independently validated and processed. This segmentation allows for more reliable detection of overlapping text by separating character identification from word boundary determination, reducing the propagation of errors that would occur in a monolithic system
Solution Approach 2:
The affinity map acts as an intermediary that reliably determines word boundaries by analyzing spatial relationships between characters. This intermediate step ensures reliable word segmentation even in complex overlapping scenarios, as it explicitly models the affinity and spatial relationships rather than attempting to directly classify words from raw input
Data Source
AI summary
Solutions for more efficient and effective optical character recognition with respect to an input text segment are disclosed. In one example, a method includes processing an input text image using a deep character overlap detection machine learning model in order to generate a character map for the input text image, an overlap map for the input text image, and an affinity map for the input text image; generating an overlap-aware word boundary recognition output based at least in part on the character map, the overlap map, and the affinity map, wherein the overlap-aware word boundary recognition output describes one or more inferred word regions of the input text image; and performing one or more prediction-based actions based at least in part on the overlap-aware word boundary recognition output.