Image Training Model Using SVD for Fast Accurate Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning methods for image processing require extensive computational resources and long training times due to the need for multilayered neural networks and iterative optimization, making it difficult to achieve highly accurate inference efficiently.
Innovation Solution
A training device that utilizes singular value decomposition (SVD) to generate a trained model by converting training images into one-dimensional signals, arranging them in matrices, and calculating right singular vectors and values, enabling rapid inference without iterative optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning is performed using multilayered neural networks with iterative optimization, then inference accuracy is improved, but training time and computational complexity increase enormously
Solution Approach 1:
The patent replaces the mechanical iterative optimization process of deep neural networks with a mathematical singular value decomposition approach. Instead of repeatedly adjusting weights through forward and backward propagation, the system performs SVD on the training image matrix to directly obtain the feature extraction function, eliminating the need for iterative mechanical optimization while maintaining high inference accuracy
Solution Approach 2:
The patent fundamentally changes the parameter representation from traditional neural network weights requiring iterative optimization to singular values and vectors obtained through direct mathematical decomposition. This parameter transformation allows the system to achieve the same feature extraction capability without the time-consuming iterative optimization process
2Measurement precision
If deep learning is performed using multilayered neural networks, then inference accuracy is improved, but computational complexity increases enormously
Solution Approach 1:
The patent substitutes the complex mechanical structure of multilayered neural networks with a streamlined mathematical approach using singular value decomposition. The system represents the training image matrix as a product of three matrices through SVD, providing a direct feature extraction mechanism that achieves high accuracy without the computational complexity of multiple neural network layers and their associated optimization algorithms
3Measurement precision
If iterative optimization processing is performed on a large number of training images, then feature vector accuracy is improved, but the amount of calculation increases enormously
Solution Approach 1:
The patent replaces the power-intensive iterative optimization process with an efficient singular value decomposition approach. Instead of repeatedly calculating gradients and updating weights across many training images, the system performs a single SVD operation on the training image matrix, directly obtaining the feature extraction function with high accuracy and minimal computational power
Solution Approach 2:
The patent performs the feature extraction function derivation in advance through singular value decomposition of the training image matrix. This preliminary action creates a ready-to-use feature extraction model that can be applied to new images without requiring additional iterative optimization, significantly reducing the computational power needed for processing
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The proposed method allows for highly accurate image inference in a significantly shorter time by bypassing the need for multilayered neural networks and iterative optimization, using SVD to generate trained models efficiently.
Implementation Method 1
a singular value decomposing unit to calculate a right singular vector and a singular value by performing singular value decomposition on the matrix generated by the matrix generating unit
Data Source
Figure 1~2
Figure 3A~3D
Figure 4A~4B
AI summary
A training device (100) includes a training image acquiring unit (110) to acquire training image information indicating a training image, a one-dimensional training image generating unit (120) to generate a plurality of one-dimensional training image signals corresponding to a plurality of pieces of the training image information by making the training image indicated by each of the plurality of pieces of the training image information into one dimension, a matrix generating unit (130) to generate a matrix in which the plurality of the one-dimensional training image signals is arranged in parallel with each other, a singular value decomposing unit (140) to calculate a right singular vector and a singular value by performing singular value decomposition on the matrix, a trained model generating unit (150) to generate a trained model that is based on the right singular vector and the singular value and outputs an inference result using, as an explanatory variable, a one-dimensional inference target image signal indicating an image obtained by making an inference target image that is an image obtained by image-capturing an inference target object into one dimension, and a trained model output unit (160) to output the trained model as trained model information are provided.