Book Page Boundary Detection via Deep Learning Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document boundary detection algorithms fail to accurately determine precise book page boundaries from a single image capture, often including background or cropping pages, and require tedious manual adjustments for two-image captures.
Innovation Solution
A deep machine learning model and image processing algorithm that computes thin and precise boundaries of individual book pages from a single image capture, using convolutional neural segmentation networks to generate pixel masks and quadrilaterals for each page, allowing for automatic detection and rendering of page boundaries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing document boundary detection algorithms are used, then the process can be automated, but the boundary detection precision is insufficient and includes background or crops pages
Solution Approach 1:
The patent divides the book spread into separate left and right pages, processing each page independently through the convolutional neural segmentation network. This segmentation allows the system to detect boundaries for each page individually, improving precision by avoiding the confusion of treating the entire spread as a single document, thus resolving the issue of including background or cropping pages.
Solution Approach 2:
The patent transforms the boundary detection approach by changing from traditional document-level boundary detection to page-level boundary detection within a spread. By adjusting the detection scope and parameters to handle individual pages separately, the system achieves more precise boundary detection that excludes background areas and maintains complete page boundaries.
2Productivity
If two-image capture method is used, then boundary detection can be performed, but manual adjustments are required which reduces productivity
Solution Approach 1:
The system performs preliminary action by automatically detecting and determining precise boundaries for both left and right pages from a single image capture using the convolutional neural segmentation network. This preliminary automatic detection eliminates the need for subsequent manual adjustments, as the system prepares the boundary information in advance, thereby improving productivity while maintaining ease of operation.
Solution Approach 2:
The scanning application performs self-service by automatically detecting page boundaries and generating the necessary boundary information without requiring user intervention or manual adjustment. The system serves itself by using the single image capture and neural network processing to automatically complete the boundary detection task that would otherwise require manual effort, thus enhancing both productivity and ease of operation.
3Productivity
If single image capture is used, then productivity increases, but existing algorithms cannot accurately determine precise boundaries
Solution Approach 1:
The patent replaces traditional mechanical or rule-based boundary detection algorithms with a convolutional neural segmentation network that uses deep learning. This substitution enables the system to achieve high measurement precision in boundary detection while maintaining the efficiency of single image capture. The neural network processes the single image to accurately determine precise boundaries for both pages, resolving the contradiction between productivity and precision.
Data Source
AI summary
Embodiments are disclosed for a precise book page boundary detection using deep machine learning model and image processing. The method may include generating, based on inputting a digital image to a convolutional neural segmentation network, a plurality of outputs comprising first output for a boundary of a left page of a spread of a physical book and second output for a boundary of a right page of the spread. A first thin pixel mask and a second thin pixel mask are determined for the left page and the right page based on the first and second outputs. A first quadrilateral and a second quadrilateral are determined based on the first and second thin pixel masks. At least one of the first or second quadrilateral is rendered in a display showing the outer boundary of the left or right page.