Deep Learning Shadow Removal for Mobile Document Scanning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Mobile phone document scanning is often hindered by shadows from the device itself or other objects, which obstructs the scanning process and reduces its effectiveness.
Innovation Solution
A device equipped with a deep learning-based shadow removal system that includes a shadow prediction model and a shadow removing model, trained using a sample library to extract and remove shadows from captured document images, utilizing a convolutional neural network architecture to improve scanning accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If mobile phone is used for document scanning, then scanning function can be realized without scanner, but shadows from mobile phone or other objects block documents and render scanning ineffective
Solution Approach 1:
The patent applies the 'Blessing in disguise' principle by using deep learning models to detect and remove shadows from document images. The shadow prediction model predicts shadow regions based on document characteristics, and the shadow removing model generates corrected images by filling in shadowed areas using learned patterns from training data. This converts the harmful shadow effect into a solvable image processing problem, enabling mobile phone scanning to function effectively despite the presence of shadows.
2Measurement precision
If deep learning-based shadow removal system is applied, then scanning accuracy and precision are improved, but device complexity increases
Solution Approach 1:
The patent applies the 'Copying' principle by using a shadow prediction model that creates a virtual representation of shadow regions based on document images. Instead of physically eliminating shadows through complex hardware modifications, the system copies the document content from non-shadowed areas and reconstructs the shadowed regions using the predicted shadow mask. This software-based approach improves scanning accuracy while avoiding the complexity of physical hardware changes.
Solution Approach 2:
The patent replaces mechanical/optical shadow removal solutions with a computational approach. Instead of using physical light sources or optical components to eliminate shadows, the system uses deep learning models to predict and remove shadows digitally. This substitution of mechanical systems with computational algorithms achieves high scanning accuracy while maintaining relatively simple device architecture.
Data Source
AI summary
A method for acquiring shadow-free images of a document for scanning or other purposes is applied in a device. The method includes training a shadow prediction model based on sample documents of a sample library and inputting a background color and a shadow mask of each of the sample documents extracted by the shadow prediction model into a predetermined shadow removing network for training, to obtain a shadow removing model. The method further includes obtaining a background color and a shadow mask of the document through the shadow prediction model and removing the shadows of the document based on the shadow removing model. The device utilizing the method is also disclosed.