Generator Network Training Using Tensor Differences for Diverse Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional generative adversarial network (GAN) models are limited in generating pictures with diverse pattern characteristics, as they primarily produce images similar to those in the original training data set, lacking the ability to create creative and new pictures.
Innovation Solution
A training method for a generator network model that involves extracting tensor matrices representing different pictures, generating new pictures based on parameter differences, and selecting new sample pictures with similarity lower than or equal to a threshold, allowing for the generation of creative and new pictures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a conventional generative adversarial network model is trained using original training data, then the generator can produce pictures with high similarity to original pictures, but the generated pictures lack pattern characteristic variety and creativity
Solution Approach 1:
The patent applies parameter changes by introducing difference values as new parameters in the training process. Specifically, difference values are calculated between original pictures and generated pictures, and these difference values are used to generate modified training data. This transforms the training approach from simply copying original pictures to actively incorporating variation through parameter modification, thereby achieving both similarity and pattern diversity
Solution Approach 2:
The patent implements preliminary action by pre-calculating difference values between original pictures and generated pictures before using them for training. These pre-computed difference values serve as additional training features that guide the generator to create pictures with both similarity to originals and sufficient pattern variety, resolving the contradiction before it affects training outcomes
2Reliability
If the generator network model is trained to generate pictures highly similar to original pictures, then the discrimination function is satisfied, but the model cannot generate creative and new pictures
Solution Approach 1:
The patent introduces difference values as an intermediary element between original pictures and generated pictures. These difference values act as a mediator that carries information about both the original picture characteristics and the desired variations. By incorporating difference values into the training process, the generator learns to produce pictures that satisfy the discrimination function while introducing creative variations through the intermediary difference value information
Solution Approach 2:
The patent performs preliminary calculation of difference values before the actual generation process. This pre-computation of difference values allows the training model to anticipate and incorporate both similarity requirements and creative variation requirements, enabling the generator to produce pictures that are both discriminatively valid and creatively novel
Data Source
AI summary
A training method of a generator network model and an electronic device for execution thereof are provided. The training method includes: extracting a first tensor matrix and a second tensor matrix, wherein the first tensor matrix and the second tensor matrix respectively represent a first picture and a second picture and individually include a plurality of first parameters and a plurality of second parameters; generating a plurality of third pictures according to a plurality of difference values between the first parameters of the first tensor matrix and the second parameters of the second tensor matrix; performing a similarity test on a plurality of original pictures and the plurality of third pictures; and adopting at least one of the third pictures whose similarity is lower than or equal to a similarity threshold as at least one new sample picture.