Discriminator Model Training via Overlapping Image Patches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in efficiently training discriminator models for accurate image classification and discrimination tasks, particularly in scenarios where data is limited or when there is a need for continuous model updates.
Innovation Solution
The proposed solution involves an electronic device that extracts overlapping image patches from training images, generates feature maps using a discriminator model, and trains the model using an objective function based on partial feature maps. This approach allows for the update of model parameters and the training of both discriminator and generator models within a GAN framework.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the discriminator model is trained using traditional methods with entire images, then the model structure is simple, but the training efficiency is low and the model performance is insufficient
Solution Approach 1:
The patent divides the training process into multiple stages: first extracting multiple patches from each training image, then training the discriminator model on these patches before combining results. This segmentation of the training data into smaller units (patches) enables more efficient processing while maintaining model performance through localized feature learning.
Solution Approach 2:
The patent introduces a new dimension to the training approach by processing images at multiple patch levels and combining results. Instead of processing entire images in a single pass, the method adds a spatial decomposition dimension, training the model on patch-level features and aggregating outcomes to achieve both efficiency and accuracy.
2Measurement precision
If overlapping image patches are extracted for training, then the training efficiency and model quality improve, but the computational complexity increases
Solution Approach 1:
The patent applies local quality by extracting multiple overlapping patches from training images, allowing the discriminator model to learn detailed local features. Each patch provides localized information that contributes to overall discrimination accuracy, with overlapping regions reinforcing important features through repeated exposure.
Solution Approach 2:
The method merges the discrimination results from multiple patches to form the final output. By combining the predictions from individual patch analyses, the system achieves comprehensive image discrimination that leverages information from all regions while managing computational load through parallel processing of patches.
3Quantity of substance
If the model is trained with limited data, then the data requirement is reduced, but the training effectiveness decreases
Solution Approach 1:
The patent segments each training image into multiple patches, effectively increasing the amount of training data derived from limited images. This segmentation allows the model to learn from more varied examples even when the total number of source images is small, improving training effectiveness without requiring large datasets.
Solution Approach 2:
The method performs preliminary feature extraction and patch creation from training images before the main training process. This preliminary action prepares enriched training data in advance, ensuring that even limited source images are fully utilized to maximize training effectiveness with minimal data requirements.
Data Source
AI summary
Training of a machine learning model is included. An electronic device includes a memory storing a machine learning model including a discriminator model, and a processor configured to extract, from a training image, a first image patch and a second image patch at least partially overlapping the first image patch, extract, from the first image patch, a first feature map based on a layer of the discriminator model, extract, from the second image patch, a second feature map based on the layer, extract a first partial feature map from a projected map that is projected based on the first feature map, and train the discriminator model based on a first objective function value generated based on a second partial feature map and the first partial feature map, wherein the second partial feature map corresponds to a part of the extracted second feature map.