Document Boundary Detection Using CNN and Image Processing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional techniques for document boundary detection in digital images struggle to accurately differentiate the document edge from background clutter, especially in cases where the document is off-center or at an angle, requiring additional manual steps and consuming significant computational resources.

Innovation Solution

A deep learning model using a convolutional neural network (CNN) in combination with image processing algorithms to identify and rate quadrilaterals approximating the document boundaries, followed by the selection of the best-rated quadrilateral to define the document edges, allowing for automatic removal of background elements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional boundary detection techniques are used, then the system is simple to implement, but the detection accuracy is poor and cannot differentiate document edges from background clutter

Engineering Contradiction:

Improveboundary detection accuracyVSAvoidsystem complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the boundary detection task into multiple components: gradient computation, non-maximum suppression, hysteresis thresholding, and quadrilateral formation. Each component processes specific aspects of edge detection separately, improving overall accuracy while maintaining manageable system complexity through modular organization of detection steps

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces intermediate structures including gradient images, thresholded edge maps, and candidate quadrilateral sets as mediators between the input image and final boundary detection. These intermediate representations facilitate accurate differentiation between document edges and background clutter by progressively refining detection candidates

Inventive Principle:

Principle #24Intermediary (Mediator)

2Measurement precision

If manual removal of background elements is performed, then the boundary detection accuracy improves, but the user interaction time and processing steps increase

Engineering Contradiction:

Improveboundary detection accuracyVSAvoiduser interaction time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system performs self-service by automatically computing gradients, applying thresholding algorithms, forming candidate quadrilaterals, and selecting the optimal document boundary without requiring user intervention. The algorithm independently differentiates document edges from background clutter through computational processing, eliminating manual removal steps

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent performs preliminary actions by pre-computing gradient images and thresholded edge maps before final boundary selection. Candidate quadrilaterals are formed and evaluated in advance, allowing the system to automatically identify and select the correct document boundary without requiring user review or manual correction

Inventive Principle:

Principle #10Preliminary action

3Extent of automation

If existing automatic boundary detection techniques are used, then user interaction is reduced, but the computational expense is high and processing speed is slow

Engineering Contradiction:

Improveautomation levelVSAvoidcomputational resource consumption

Core Design Contradiction:

Extent of automationVSUse of energy by moving object

Solution Approach 1:

The patent applies partial action by computing gradients and thresholding only in regions where document boundaries are likely to occur, rather than processing the entire image uniformly. Candidate quadrilaterals are formed from significant edges only, reducing computational expense while maintaining high automation level for boundary detection

Inventive Principle:

Principle #16Partial or excessive action

Solution Approach 2:

The detection process is segmented into distinct computational stages: gradient computation, thresholding, edge tracking, and quadrilateral formation. Each stage processes only relevant data from the previous stage, reducing overall computational resource consumption while maintaining full automation throughout the pipeline

Inventive Principle:

Principle #1Segmentation

4Adaptability or versatility

If conventional techniques are used, then the system is computationally efficient, but it cannot handle off-center or angled documents effectively

Engineering Contradiction:

Improvedocument orientation handlingVSAvoidalgorithm complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent handles asymmetric document orientations by forming quadrilateral candidates from detected edges without assuming symmetric or centered document placement. The algorithm evaluates quadrilaterals based on their geometric properties and edge consistency, allowing effective detection of off-center and angled documents through asymmetric boundary formation

Inventive Principle:

Principle #4Asymmetry

Data Source

PatentUS10970847B2Document boundary detection using deep learning model and image processing algorithms

Publication Date: 2021.04.06 ADOBE INC

AI summary

Techniques are disclosed for document boundary detection (BD) from an input image using a combination of deep learning model and image processing algorithms. Quadrilaterals approximating the document boundaries in the input image are determined and rated separately using both these approaches: deep leaning using convolutional neural network (CNN) and heuristics using image processing algorithms. Thereafter, the best rated quadrilateral is selected from the quadrilaterals obtained from both the approaches.