Blur Classification Probability Matrix for Digital Image Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating the blur state of digital image pixels are limited by inaccurate models of the physical blur process and are influenced by image content, such as shadows and smooth backgrounds, and struggle to distinguish between different blur kernels and sharp edges.
Innovation Solution
A system and method that computes the probability of pixel blur using a two-dimensional blur kernel, generating a blur classification probability matrix and performing binary segmentation into sharp and blurry regions through pre-processing and frequency power spectrum analysis, allowing for coherent mapping and image editing operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If inverse diffusion technique is used to estimate blur, then blur estimation can be performed, but the model assumes Gaussian blur kernel which is not accurate and image content unduly affects the blur estimate
Solution Approach 1:
The patent changes the blur kernel model from a fixed Gaussian function to a two-dimensional boxcar (rectangular) kernel, which more accurately represents the physical optical transfer function of camera lenses. This parameter change fundamentally improves the accuracy of blur estimation by using a more appropriate mathematical model that matches actual lens behavior.
Solution Approach 2:
The patent creates a synthetic blurred image by convolving the original sharp image with a boxcar kernel, then compares this synthetic blur with the actual observed blur. This copying approach allows the system to isolate and measure the true blur effect independent of image content variations, solving the problem of content-dependent estimation errors.
2Adaptability or versatility
If power spectrum of local windows is used to estimate blur, then blur estimation can be performed, but it is limited to one-dimensional blur kernels and uses sharp edged windows that cannot distinguish window edges from image edges
Solution Approach 1:
The patent extends the blur kernel from one-dimensional (1D) to two-dimensional (2D), allowing the system to model realistic circular or rectangular point spread functions that correspond to actual lens optics. This dimensional change enables accurate modeling of radial blur patterns and provides superior edge detection capability compared to 1D approaches.
Solution Approach 2:
The patent applies different window functions (e.g., Gaussian, Hamming) with smooth edges rather than sharp-edged windows. This local quality improvement allows the window edges to be distinguishable from actual image edges, as the smooth window transitions create characteristic frequency signatures that preserve edge information while reducing artificial artifacts.
3Ease of manufacture
If single most likely blur kernel is assumed, then segmentation can be performed, but it cannot distinguish between different blur kernels and assumes uniform blur across regions
Solution Approach 1:
The patent segments the image into multiple regions and estimates a separate blur kernel for each region, allowing different parts of the image to have different blur characteristics. This segmentation approach maintains computational feasibility while enabling the system to distinguish between multiple blur kernels and model non-uniform blur patterns throughout the image.
Solution Approach 2:
The patent makes the blur kernel model dynamic by allowing different regions of the image to have different blur kernel parameters (size, shape, orientation). This dynamic approach enables the system to adapt to varying focus conditions, motion blur directions, and optical characteristics across different parts of the image, rather than assuming a single static blur model.
Data Source
AI summary
A blur classification module may compute the probability that a given pixel in a digital image was blurred using a given two-dimensional blur kernel, and may store the computed probability in a blur classification probability matrix that stores probability values for all combinations of image pixels and the blur kernels in a set of likely blur kernels. Computing these probabilities may include computing a frequency power spectrum for windows into the digital image and/or for the likely blur kernels. The blur classification module may generate a coherent mapping between pixels of the digital image and respective blur states, and/or may perform a segmentation of the image into blurry and sharp regions, dependent on values stored in the matrix. Input image data may be pre-processed. Blur classification results may be employed in image editing operations to automatically target image subjects or background regions, or to estimate the depth of image elements.