Rotational Invariant Neural Network Training via Symmetric Kernels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep convolutional neural networks are computationally expensive due to their complexity, especially with large input sizes, and their output depends on the orientation of the input image, leading to varying results when applied to differently oriented images.
Innovation Solution
A method and system for training neural networks to be symmetric, ensuring that the output is invariant to image rotations and orientations by using symmetric kernels and loss functions that minimize differences across different orientations, thereby reducing computational complexity and achieving consistent performance across image orientations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep convolutional neural networks are used for image processing, then image processing performance is improved, but computational complexity and memory requirements increase significantly
Solution Approach 1:
The patent applies asymmetry by using asymmetric convolutional kernels (e.g., 3x1 and 1x3 kernels instead of symmetric 3x3 kernels) to reduce the number of parameters and computations while maintaining image processing performance. This asymmetric kernel approach reduces computational complexity and memory requirements by eliminating redundant calculations in different spatial directions.
2Speed
If standard convolutional kernels are used, then computational speed is maintained, but output varies with image orientation
Solution Approach 1:
The patent implements universality by designing convolutional kernels that perform multiple functions: they maintain computational efficiency like standard kernels while simultaneously providing orientation-invariant feature extraction. The asymmetric kernel design with specific weight sharing patterns enables the network to process images of any orientation consistently, making the kernel universally applicable across different image orientations without requiring separate kernels for each orientation.
3Reliability
If symmetric kernels are trained to achieve orientation invariance, then output consistency across orientations is improved, but training complexity increases
Solution Approach 1:
The patent applies local quality by imposing symmetry constraints specifically on the convolutional kernel weights rather than the entire network. This localized approach to symmetry enforcement at the kernel level achieves orientation-invariant outputs while minimizing the impact on overall training complexity. The symmetry constraint is applied locally to weight parameters, allowing the rest of the network to train normally with standard optimization algorithms.
Data Source
AI summary
A method and apparatus for providing a rotational invariant neural network is herein disclosed. According to one embodiment, a method includes receiving a first input of an image in a first orientation and training a kernel to be symmetric such that an output corresponding to the first input is the same as an output corresponding to a second input of the image in a second orientation.