Neural Network Manuscript Character Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems struggle to accurately extract manuscript characters from scanned images when the figure information is not preregistered, limiting their ability to process images with unregistered content.
Innovation Solution
An image-processing system and method that uses a neural network trained with foreground and background sample images to identify and remove manuscript characters from any image, regardless of preregistered figure information, by generating training data through image composition and deep training with a multilayer neural network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If preregistered figure information is used to extract manuscript characters, then extraction accuracy is improved for registered figures, but the system cannot extract manuscript characters from images containing unregistered figures
Solution Approach 1:
The system performs preliminary training by composing training images that include both registered and unregistered figure information before actual manuscript character extraction. This preliminary preparation enables the extraction algorithm to handle diverse figure types without requiring preregistration, resolving the contradiction between extraction accuracy and adaptability to unregistered figures
Solution Approach 2:
The system develops a universal manuscript character extraction capability that works with both registered and unregistered figures through image composition training. The trained model becomes multi-functional, able to extract manuscript characters regardless of whether the underlying figure information is preregistered, thus achieving versatility without sacrificing accuracy
2Productivity
If traditional extraction methods are used, then processing speed is maintained, but the system cannot accurately extract manuscript characters from images with unregistered content
Solution Approach 1:
The system performs image composition and neural network training in advance as a preliminary action. This pre-processing creates an optimized extraction model that achieves both high accuracy for unregistered figures and maintains efficient processing speed during actual operation, eliminating the trade-off between speed and accuracy
3Device complexity
If the system is designed to handle only registered figures, then device complexity is reduced, but the system lacks capability to process diverse document types
Solution Approach 1:
The system uses image composition to create synthetic training samples that replicate various document types and figure combinations. By copying and combining existing registered figure images to create diverse training scenarios, the system achieves high adaptability to unregistered figures without requiring complex additional hardware or structure
Solution Approach 2:
The system achieves versatility through parameter changes in the training phase, varying figure types, combinations, and configurations in composed training images. This allows the neural network to learn diverse patterns while maintaining a relatively simple system structure during actual extraction operations
Data Source
Figure 1
Figure 2A
Figure 2B~2C
AI summary
An image-processing system (100) that can perform predetermined image processing on a manuscript character in a captured image is provided. A method for image processing includes training a neural network using composite image data including a background image and a manuscript image and correct image data corresponding to the composite image data, obtaining a captured image of an original that contains a manuscript character, and performing predetermined image processing on the captured image using the neural network.