Handwritten Text Image Generation for Deep Learning Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning models face challenges in effectively recognizing handwritten text due to the variability in handwriting styles and the lack of sufficient image samples for training.
Innovation Solution
A method is provided for generating image samples with annotation boxes, which includes creating handwritten text images from sample images and combining these with background images to produce target sample images. This approach enables the generation of a large number of sample images for training detection models, improving their efficiency in recognizing handwritten text.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are used for text recognition, then recognition accuracy can be improved, but the lack of sufficient training samples limits model performance
Solution Approach 1:
The patent uses template images to generate multiple handwritten text sample images through copying and transformation. A single template image is copied and transformed with different backgrounds, noise levels, and styles to create diverse training samples, effectively multiplying the available training data from limited original samples
Solution Approach 2:
The patent changes various parameters of the training samples including background images, noise levels, text styles, and image transformations. By modifying these parameters, the system generates diverse sample images that maintain the core text information while varying visual characteristics, thereby increasing sample quantity and diversity
2Productivity
If manual recognition is used for handwritten text, then labor costs are high, but automated recognition lacks sufficient training data
Solution Approach 1:
The patent automatically generates large numbers of training samples by copying template images and applying various transformations. This automated copying process creates sufficient training data to enable effective deep learning model training, replacing the need for manual recognition while providing adequate training materials
Solution Approach 2:
The patent performs preliminary actions by pre-processing template images and pre-generating various transformed versions before actual recognition tasks. These pre-generated samples serve as ready-to-use training data, eliminating the need for manual sample collection and preparation during deployment
Data Source
AI summary
A method of generating an image sample, which relates to a field of an artificial intelligence technology, in particular to fields of a deep learning technology and a computer vision technology. The method includes: generating a handwritten text image according to at least one handwritten sample image; and generating a target sample image with an annotation box according to the handwritten text image and a background image, where the annotation box is used to represent a region in which the handwritten text image is located in the background image. The present disclosure further provides a method of recognizing a text, an electronic device and a storage medium.