Document Posture Alignment Using CNN Inference for eKYC Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies, such as those described in WO 2020/008628 A1, require extensive feature point extraction, leading to increased processing loads on computers, especially when handling continuous document captures, which can hinder efficient identity verification processes like eKYC.
Innovation Solution
An image processing system that includes a learning terminal and server to train a learning model to correct the posture of captured documents by using a convolutional neural network to align feature points between training and reference images, reducing processing load through efficient image alignment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature point extraction is performed to correct document posture, then document alignment accuracy is improved, but processing load on computer increases
Solution Approach 1:
The patent applies preliminary action by pre-training a learning model with ground truth data that encodes document posture correction knowledge. During actual processing, the model directly predicts correction parameters without performing computationally intensive feature point extraction, thus reducing real-time processing load while maintaining alignment accuracy.
Solution Approach 2:
The patent replaces the mechanical feature point extraction and matching system with a neural network-based prediction system. The learning model substitutes traditional image processing algorithms, transforming the mechanical computation of feature points into a streamlined neural network inference process that achieves the same alignment goal with lower computational burden.
2Measurement precision
If feature point extraction is performed on continuously captured document images, then document posture correction is achieved, but processing time increases
Solution Approach 1:
The learning model is pre-trained offline with extensive feature point data and ground truth corrections. This preliminary training phase shifts the computational burden to an offline setting, enabling rapid real-time inference when processing continuously captured document images, thus improving processing speed without sacrificing correction accuracy.
Solution Approach 2:
Instead of performing complete feature point extraction and matching for every captured image, the system uses the learning model to predict only the essential correction parameters. This partial action approach processes only the critical information needed for posture correction, significantly reducing processing time while maintaining sufficient accuracy for continuous capture scenarios.
Data Source
AI summary
Provided is an image processing system including at least one processor configured to: acquire training data including, as an input portion, a training target image in which a training target document is shown and a training reference image in which a training reference document is shown and including, as a ground truth portion, ground truth information for processing the training target image so that a training target posture of the training target document in the training target image matches a training reference posture of the training reference document in the training reference image; and train, based on the training data, a learning model for image processing so that the ground truth information is output when the training target image and the training reference image are input.