Large Image Matting via Sub-Image Segmentation and Alpha Mask Stitching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image matting methods, such as the closed form method, face memory constraints when dealing with large images, leading to memory shortages and inability to process images larger than 8 megapixels due to high memory consumption, resulting in resolution loss and unbalanced alpha regions.
Innovation Solution
Divide large images into smaller sub-images, apply the modified closed form method to each sub-image using alpha values from a reduced original image as constraints, and combine the alpha masks to form a high-resolution alpha mask without interpolation, allowing for practical usage on systems with limited memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the closed form method is applied to large images, then matting quality is improved, but memory consumption increases beyond available capacity
Solution Approach 1:
The patent divides the large image into multiple smaller sub-images (tiles) that can be processed independently within available memory constraints. Each sub-image is processed separately using the closed form method, and the resulting alpha masks are stitched together to form the final high-resolution matting result. This segmentation allows the application of memory-intensive closed form method to large images by reducing the memory footprint of individual processing units.
2Quantity of substance
If image resolution is reduced to fit memory constraints, then memory consumption is decreased, but alpha mask resolution is lost
Solution Approach 1:
Instead of reducing the entire image resolution, the patent segments the large high-resolution image into smaller sub-images that individually fit within memory constraints. This allows processing at full resolution without the need for downsampling, thereby preserving alpha mask quality while accommodating memory limitations.
Solution Approach 2:
The patent processes the image in spatial segments (dividing the 2D image space into multiple tiles) rather than reducing the resolution dimension. This dimensional approach to problem-solving maintains full resolution in the processed dimensions while managing memory through spatial partitioning.
3Quantity of substance
If the image is divided into sub-images, then memory consumption is reduced, but processing complexity increases
Solution Approach 1:
The image is divided into overlapping sub-images (tiles) that can be processed independently. The overlap region between adjacent tiles is used to ensure continuity and smooth transitions in the final stitched alpha mask. This segmentation strategy balances memory reduction with manageable processing complexity by creating independent, reusable processing units.
Solution Approach 2:
The patent pre-calculates and stores the alpha masks for the overlapping boundary regions of each sub-image. These pre-computed boundary alpha values are then used as constraints when processing adjacent sub-images, ensuring continuity without requiring complex iterative adjustments during the stitching phase.
4Quantity of substance
If existing downscaling methods are used, then memory constraints are addressed, but alpha regions become unbalanced
Solution Approach 1:
By processing overlapping sub-images independently and stitching them together, the patent maintains the original color and alpha distribution characteristics of each region. This avoids the global reprocessing and interpolation operations in downscaling methods that can distort alpha region balances, while still addressing memory constraints through spatial segmentation.
Data Source
AI summary
A method and system for large image matting, where alpha mask αR for a large image is built with the help of alpha mask αR found for the reduced image. In the current context, a large image is an image for which data structures required by the closed form method do not fit available memory. To overcome memory limitation the large image is divided into a plurality of smaller size sub-images. To find an alpha mask for every sub-image a sparse linear system is solved where the pixel values sampled from reduced image alpha mask αR are used as constraints on a sparse linear system.