Non-Square Convolution Filters for Balanced Pixel Weighting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional image processing circuits using square convolution filters face imbalances in pixel weighting due to varying distances from the central pixel, leading to difficulties in capturing image characteristics effectively.

Innovation Solution

The use of non-square convolution filters, such as rhombus, dilated-rhombus, cross, dilated-cross, X-shape, and dilated-X-shape filters, which provide balanced pixel spacing to enhance image feature fetching efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If square convolution filters are used, then the computation is intuitive and convenient, but the pixel weighting becomes unbalanced due to varying distances from the central pixel

Engineering Contradiction:
Improvecomputation convenienceVSAvoidpixel weighting balance
Core Design Contradiction:
Ease of operationVSManufacturing precision

Solution Approach 1:

The patent applies asymmetry by transitioning from square convolution filters to non-square convolution filters (such as rectangular filters with different height and width). This asymmetric design allows the filter to cover a different number of pixels in vertical and horizontal directions, thereby balancing the L1 norm distances from the central pixel across all covered pixels. The asymmetric filter shape enables equal weighting of pixels at the same Manhattan distance from the center, resolving the weighting imbalance issue while maintaining computational efficiency.

Inventive Principle:
Principle #4Asymmetry

2Device complexity

If square convolution filters are used, then the filter structure is simple and uniform, but the image feature capture effectiveness is reduced due to unbalanced pixel contributions

Engineering Contradiction:
Improvefilter structure simplicityVSAvoidimage feature capture accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent employs asymmetric convolution filter structures (rectangular filters with unequal height and width) to balance the contributions of pixels at different positions. This asymmetric design ensures that all pixels covered by the filter have equal L1 norm distances from the central pixel, leading to balanced weighting and improved image feature capture effectiveness. The filter maintains structural simplicity while achieving better feature extraction through its non-square geometry.

Inventive Principle:
Principle #4Asymmetry

3Manufacturing precision

If non-square convolution filters are used, then the pixel weighting balance is improved, but the filter design becomes more complex

Engineering Contradiction:
Improvepixel weighting balanceVSAvoidfilter design complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by modifying the dimensional parameters of the convolution filter from equal dimensions (square) to unequal dimensions (rectangular). By changing the height and width parameters independently, the filter achieves balanced L1 norm distances for all covered pixels. This parameter adjustment approach maintains computational efficiency while achieving balanced pixel weighting, with the added benefit of enhanced image feature capture capabilities.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentUS11157769B2Image processing circuit and associated image processing method
Publication Date: 2021.10.26 REALTEK SEMICON CORP
  • US11157769B2 patent drawing
  • US11157769B2 patent drawing
  • US11157769B2 patent drawing

AI summary

An image processing circuit includes a receiving circuit, a feature fetching module and a decision circuit. In the operations of the image processing circuit, the receiving circuit is configured to receive image data. The feature fetching module is configured to use a multi-topological-convolutional network to fetch the features of the image data, to generate a plurality of image features determined by the characteristics and weights of the convolution filter, where the image features may be smooth features or edge features. In the present invention, the convolution filters used by the feature fetching module are not limited by a square convention filter, and the convolution filters may include the multiple topological convolutional network having non-square convolution filters. By using the multiple topological convolutional network of the present invention, the feature fetching module can fetch the rich image features for identifying the contents of the image data.