Adaptive Image Binarization Using Multi-Scale Window Variance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image binarization methods face challenges with a single window size failing to account for local variations in foreground and background intensity, leading to false classifications and computational complexity issues, especially on devices with limited resources.
Innovation Solution
The method employs multiple windows of varying sizes surrounding each pixel, using variance instead of standard deviation to determine binarization thresholds, and selects appropriate threshold generation functions based on variances to adaptively compute bi-level pixel values without requiring square root operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single window size is used for the whole image, then the computation is simple, but it cannot account for local variations in foreground and background intensity leading to false classifications
Solution Approach 1:
The image is divided into multiple windows of different sizes centered at each pixel. Instead of using a single window size for the entire image, the method segments the analysis into multiple scale levels, allowing local variations to be captured while maintaining computational efficiency through the use of variance instead of standard deviation.
Solution Approach 2:
The method applies different window sizes locally depending on the pixel position and local image characteristics. By evaluating multiple windows of varying sizes at each pixel location, the binarization process adapts to local variations in foreground and background intensity, improving classification accuracy without requiring a single fixed window size for the entire image.
2Measurement precision
If standard deviation is used to determine thresholds, then the binarization can adapt to local variations, but the computational complexity increases due to square root operations
Solution Approach 1:
The method changes the statistical parameter from standard deviation to variance for threshold determination. By using variance instead of standard deviation, the computation avoids square root operations while still capturing the necessary local variation information. This parameter substitution maintains threshold determination accuracy while significantly improving computation speed, especially on devices with limited resources.
3Productivity
If an overly-large window size is used, then computation is faster, but local variations in intensity are missed leading to false classifications
Solution Approach 1:
The analysis is segmented into multiple window sizes, allowing the method to capture both local variations through smaller windows and maintain computational efficiency through the variance-based approach. The multi-scale window segmentation ensures that no single window size dominates, balancing speed and accuracy.
Solution Approach 2:
The method adds the dimension of multiple window sizes instead of relying on a single window size. By evaluating variance across different window size dimensions, the method captures local variations that would be missed by a single large window while avoiding the computational burden of exhaustively testing all possible window sizes through the efficient variance calculation.
4Measurement precision
If a window size that is too small is used, then local detail is preserved, but the threshold computation becomes unreliable due to insufficient pixels
Solution Approach 1:
The method segments the window size evaluation into multiple scales, where small windows capture local detail and larger windows provide sufficient pixel samples for reliable variance computation. By combining results across multiple window size segments, the method ensures both local detail preservation and threshold computation reliability.
Solution Approach 2:
The method merges the results from multiple window sizes to determine the final binarization threshold. By combining information from small windows (which preserve local detail) and larger windows (which provide reliable statistical samples), the method achieves both local detail preservation and threshold computation reliability that neither window size could achieve alone.
Data Source
AI summary
Bi-level pixel values are generated from a set of input pixel values corresponding to an image. Various described methods and apparatus are well suited for applications with limited computational capability and/or limited available resources to be used for performing image processing. Corresponding to an individual input pixel being processed, a plurality of windows including the pixel are evaluated to determine statistics including a variance for each window. Based upon the determined variances, one of a plurality of binarization threshold generation functions is selected. A binarization threshold for the input pixel is determined using the selected binarization threshold generation function. A bi-level pixel value is generated based on a comparison of the input pixel value to the generated binarization threshold. In various embodiments, the binarization threshold determination functions use non-zero integer powers of one or more variances, and intentionally avoid performing a square root operation, thus limiting computational complexity.