Inverse Kernel Defocus Deblurring for Spatially Varying Blur
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deblurring methods struggle to accurately remove defocus blur due to spatially varying blur sizes and shapes, leading to irreversible information loss and limited performance, especially in deep learning-based approaches that require significant computation.
Innovation Solution
The method employs an inverse kernel-based approach, utilizing a shared reference blur shape and pixel-wise blur size estimation, combined with atrous convolutional layers and attention modules to reduce blur complexity and maintain effective deblurring performance with a smaller network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional two-step approaches assume isotropic kernel shapes (disc or Gaussian) and estimate only blur size, then computational complexity is reduced, but deblurring performance is limited due to error in blur size estimation and deviation from actual blur models
Solution Approach 1:
The patent applies local quality by estimating blur size separately for each pixel position while maintaining a common reference blur shape. This allows the blur estimation to adapt to spatial variations in blur size without requiring complex per-pixel kernel estimation, thus reducing computational complexity while improving deblurring performance compared to global isotropic assumptions.
Solution Approach 2:
The patent changes the parameter representation from assuming fixed isotropic kernel shapes to using a common reference blur shape with pixel-wise blur size parameters. This parameterization approach allows more flexible modeling of actual blur characteristics while maintaining computational efficiency through shared shape representation across all pixels.
2Manufacturing precision
If deep learning-based deblurring methods are used, then deblurring performance can be improved, but computational requirements increase significantly making them unsuitable for mobile devices
Solution Approach 1:
The patent segments the deblurring task into two parts: (1) estimating a common reference blur shape that applies to all pixels, and (2) estimating pixel-wise blur size parameters. This segmentation avoids the need for heavy deep learning models while achieving comparable performance by focusing computational resources on the most critical varying parameter (blur size) rather than modeling entire kernels for each pixel.
Solution Approach 2:
The common reference blur shape serves as a universal model for all pixels in the image, providing a shared structural foundation that reduces the number of parameters to be estimated. This universal representation enables efficient processing on mobile devices while maintaining deblurring performance through pixel-adaptive blur size adjustment.
3Manufacturing precision
If separate inverse kernels are generated for each pixel position to handle spatially varying blur sizes, then deblurring accuracy is improved, but the number of parameters and computational complexity increase significantly
Solution Approach 1:
The patent achieves local adaptation by estimating only blur size parameters for each pixel position while sharing a common reference blur shape across all pixels. This approach provides pixel-level customization where needed (blur size variation) without the full complexity of separate kernel estimation, significantly reducing the number of parameters while maintaining deblurring accuracy.
Solution Approach 2:
Instead of directly estimating complex per-pixel blur kernels, the patent inverts the approach by first establishing a common reference shape and then estimating only the size scaling parameters. This inverted parameterization strategy reduces the dimensionality of the estimation problem from full kernel estimation to simple size parameter estimation, achieving accuracy with fewer parameters.
Data Source
AI summary
The present disclosure provides a method of effectively deblurring a defocus blur in an input image based on an inverse kernel. The defocus deblurring method includes: generating, by an encoder network, an input feature map by encoding the input image; filtering, by an atrous convolution network including a plurality of atrous convolutional layers arranged in parallel, the input feature map to generate an output feature map having reduced blur component; and generating, by a decoder network, an output image having reduced blur from the output feature map with the reduced blur component and the input image.