OCR Preprocessing via Character Pixel Height Estimation and Cycle GAN
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Optical character recognizer engines fail to accurately extract text from images with distortions such as noise, bad lighting, and warping, leading to decreased recognition rates due to inadequate preprocessing methods that can either remove essential character details or fail to correct distortions effectively.
Innovation Solution
A text extraction method that calculates the estimated character pixel height, scales the image, removes distortions using a neural network trained by a cycle generative adversarial network, and postprocesses the extracted characters using a Levenshtein automaton model and deep learning language model to improve OCR accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If noise removal and erosion are applied to thicken character strokes, then character recognition accuracy is improved, but character details may be lost or characters may become unrecognizable
Solution Approach 1:
The patent replaces traditional mechanical preprocessing operations (noise removal filters, erosion morphological operations) with a deep learning-based neural network that learns optimal preprocessing transformations. This substitution allows the system to achieve noise removal and character thickening while preserving essential character details, as the neural network can distinguish between actual character features and noise patterns through training on distorted and clean image pairs.
2Reliability
If aggressive preprocessing is applied to remove noise, then recognition rate improves, but light characters may be removed as noise
Solution Approach 1:
The patent replaces aggressive mechanical noise removal operations with a trained neural network that can differentiate between noise and light character elements. The network learns during training which patterns represent actual characters versus noise, allowing it to preserve light characters while removing noise, thereby maintaining both high recognition rates and character integrity.
Solution Approach 2:
The patent transforms the preprocessing approach by changing from fixed-parameter mechanical operations to adaptive parameter adjustments through neural network learning. The network dynamically adjusts preprocessing intensity and characteristics based on the specific input image characteristics, allowing it to handle light characters appropriately without removing them as noise.
3Productivity
If traditional preprocessing methods are used, then processing speed is maintained, but distortion correction is insufficient
Solution Approach 1:
The patent replaces traditional mechanical distortion correction methods with a deep learning-based neural network that can perform complex nonlinear transformations. This substitution enables superior distortion correction quality while maintaining processing speed, as the trained network can apply complex corrections in a single pass without requiring multiple iterative mechanical operations.
Data Source
AI summary
A text extraction computing method that comprises calculating an estimated character pixel height of text from a digital image. The method may scale the digital image using the estimated character pixel height and a preferred character pixel height. The method may binarizes the digital image. The method may remove distortions using a neural network trained by a cycle GAN on a set of source text images and a set of clean text images. The set of source text images and clean text images are unpaired. The source text images may be distorted images of text. Calculating the estimated character pixel height may include summarizing the rows of pixels into a horizontal projection, and determining a line-repetition period from the projection, and quantifying the portion of the line-repetition period that corresponds to the text as the estimated character pixel height. The method may extract characters from the digital image using OCR.