Sub-region LBP Feature Extraction Reducing Noise and Load

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing local binary pattern (LBP) techniques face a trade-off between increasing the region size for feature extraction, which doubles the number of bits and thus increases processing load, and maintaining detection accuracy, as wider regions lead to higher dimensionality histograms and increased noise sensitivity.

Innovation Solution

The proposed solution involves setting sub-regions constituted by multiple pixels and calculating representative values for these sub-regions, generating local binary patterns based on differences with the pixel of interest, and using these representative values to reduce the number of bits and noise, thereby allowing more pixels to be used in feature extraction while limiting processing load and accuracy decrease.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Area of stationary object

If a wider region (e.g., 5 pixels × 5 pixels) is targeted for arithmetic operation of local binary pattern, then more pixels can be used for feature extraction, but the number of bits of local binary pattern increases (from 8 to 16), leading to increased processing load and histogram dimensionality

Engineering Contradiction:
Improveregion size for feature extractionVSAvoidprocessing load
Core Design Contradiction:
Area of stationary objectVSDevice complexity

Solution Approach 1:

The patent divides the neighboring region into multiple sub-regions (e.g., four 2×2 sub-regions within a 4×4 region). Instead of computing LBP using all pixels in the entire neighboring region, the patent computes LBP by comparing the pixel of interest with representative values (e.g., average intensity) of each sub-region. This segmentation reduces the number of bits required while still capturing broader spatial information.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different processing strategies to different parts of the image by dividing the neighboring region into sub-regions. Each sub-region is processed independently to generate a representative value, allowing the system to capture local variations while reducing overall computational complexity. This enables selective focus on important local patterns without processing every pixel uniformly.

Inventive Principle:
Principle #3Local quality

2Adaptability or versatility

If a wider region is targeted for arithmetic operation, then co-occurrence of features in wider range can be utilized, but the number of dimensions of histogram increases, leading to increased processing load in generating classifier or detecting object

Engineering Contradiction:
Improvefeature co-occurrence rangeVSAvoidclassifier generation speed
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent segments the neighboring region into multiple sub-regions and computes LBP features by comparing the pixel of interest with representative values of these sub-regions. This approach enables the capture of feature co-occurrence over a wider spatial range while maintaining a manageable number of histogram dimensions, thus preserving adaptability without sacrificing productivity.

Inventive Principle:
Principle #1Segmentation

3Quantity of substance

If the number of bits of local binary pattern increases, then more pixels are used in feature extraction, but the number of dimensions of histogram increases, leading to increased noise sensitivity

Engineering Contradiction:
Improvenumber of pixels usedVSAvoidnoise sensitivity
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent divides the neighboring region into sub-regions and uses representative values (such as average intensity) of these sub-regions for LBP computation. This segmentation approach allows more pixels to be utilized in feature extraction while reducing noise sensitivity by aggregating pixel values within sub-regions before comparison, thereby improving reliability.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP2743888B1Feature extraction device, feature extraction program, and image processing device
Publication Date: 2023.10.25 PANASONIC INTELLECTUAL PROPERTY MANAGEMENT CO LTD
  • EP2743888B1 patent drawingFigure 1
  • EP2743888B1 patent drawingFigure 2
  • EP2743888B1 patent drawingFigure 3

AI summary

Provided is a feature extraction device including a sub region setting unit (433) which sets a plurality of sub regions with respect to a pixel of interest, and a binary pattern generation unit (434) which generates a local binary pattern indicating pixel value comparison results relative to each sub region with respect to each pixel of interest. The sub region setting unit (433) sets (436) a region constituted by a plurality of pixels including a pixel separated from the pixel of interest as a sub region, and the binary pattern generation unit (434) calculates (437) a representative value for each sub region and generates (439) a local binary pattern indicating whether or not the difference (438) between the representative value and the value of the pixel of interest is equal to or larger than a predetermined threshold value.