Handwriting Image Synthesis for Noise-Robust Character Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods struggle to effectively learn and extract handwritten characters due to fixed scanning noise positions, leading to reduced accuracy in character recognition.
Innovation Solution
Generate learning data by synthesizing images with randomly positioned scanning noise to create robust neural networks capable of accurately extracting handwritten characters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning data is created by processing images using scan images as disclosed in U.S. Pat. No. 10,546,217, then the scanning noises are fixed and they appear at the same positions, but performing robust learning against the influence of the noises may be difficult
Solution Approach 1:
The patent applies dynamics by making the noise positions variable rather than fixed. Specifically, it superimposes scan images onto the background image at different positions for each training sample, creating dynamic variation in noise locations. This enables the learning model to adapt to various noise positions and improves robustness against scanning noises.
Solution Approach 2:
The patent changes the positional parameters of noise superposition. By varying the superposition position of scan images across different training samples, it introduces parameter diversity that helps the model learn to handle noises at various locations, thereby improving robustness while maintaining adaptability.
2Measurement precision
If the learning is performed using learning data generated according to U.S. Pat. No. 10,546,217, then it is not possible to sufficiently learn the influence of noises caused by scanning, but there is a possibility that the extraction accuracy of handwritten characters deteriorates due to the influence of noises
Solution Approach 1:
The patent applies preliminary action by pre-training the learning model with scan images that contain actual scanning noises. By incorporating real noise patterns from scan images into the training data before final training, the model learns to recognize and ignore noise patterns in advance, thereby improving both noise robustness and character extraction accuracy.
Solution Approach 2:
The patent converts the harmful effect of scanning noises into a beneficial training resource. By using scan images containing noises as part of the training data and superimposing them at various positions, the model learns to tolerate and filter out these noises, transforming what would be detrimental interference into a valuable learning opportunity that improves extraction accuracy.
Data Source
AI summary
The information processing apparatus according to the present disclosure synthesizes a handwriting image with a noise image to generate a synthesized image, generates a correct label indicative of handwriting pixels from the handwriting image, and applies the synthesized image and the correct label as learning data to generate a learning model.