Layered Text Extraction for Accurate Identification in Busy Backgrounds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional text identification applications struggle to accurately identify text layered over objects in busy backgrounds due to the use of entire layered digital images as input, leading to inefficiencies and inaccuracies, and require significant data annotation efforts.
Innovation Solution
Generate a text channel image by isolating the text layer from other layers, presenting it on a white background, and use a page decomposition model trained with machine learning to enhance text identification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the entire layered digital image is used as input for text identification, then all visual information is preserved, but text identification accuracy deteriorates due to visual background noise interfering with text recognition
Solution Approach 1:
The patent segments the layered digital image into multiple channels including a text channel, image channel, and vector channel. The text channel is extracted and processed separately to eliminate interference from background images and vector elements, thereby improving text identification accuracy while maintaining the integrity of the original multi-layer structure
Solution Approach 2:
The patent extracts the text channel from the layered digital image by isolating text layer data from image and vector layers. This extraction removes harmful visual background noise that would otherwise interfere with text recognition, allowing the text identification model to process only relevant text information
2Ease of manufacture
If conventional text identification methods are used on layered digital content, then implementation is straightforward, but text identification accuracy deteriorates in complex backgrounds with significant objects
Solution Approach 1:
The patent divides the layered digital content into separate channels (text, image, vector) and processes them independently. This segmentation allows conventional identification methods to be applied to the isolated text channel where they remain simple to implement, while avoiding the accuracy problems that arise when processing the entire complex image
3Reliability
If data annotation is performed on entire layered digital images, then comprehensive training data is obtained, but the amount of annotation work increases significantly
Solution Approach 1:
The patent extracts the text channel from layered digital images before performing data annotation. This extraction reduces the complexity of the annotation task by removing background images and vector elements, thereby maintaining training data quality while significantly reducing the time and effort required for annotation
Data Source
AI summary
Techniques for text identification in layered digital content are described. In an implementation, an item of digital content is received including a plurality of layers. A text layer is identified in the item of digital content from the plurality of layers. A text channel image is generated by isolating the text layer from the plurality of layers. A text identification is then generated based on the text channel image using a page decomposition model, the page decomposition model trained using machine learning.