Image Learning Model for Synthetic Face Generation and Noise Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image recognition technologies face challenges in generating images at desired angles and improving recognition rates due to insufficient learning images, and are limited in generating new faces through combining patches from training sets.
Innovation Solution
The method employs an image learning model, specifically a neural network with connection weights, to generate synthetic images by combining feature values from input images and to produce images at different angles, while also eliminating noise from images using separate learning models for different patch sizes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional patch combining methods are used to generate new faces, then some face generation is enabled, but the generation capability is limited and cannot produce faces at desired angles
Solution Approach 1:
The patent transforms discrete patch combinations into continuous parameter-based transformations. By representing faces in a latent space and using linear combinations with controlled parameters (alpha, beta coefficients), the system can generate faces at any angle or intermediate state, not just discrete patches from the training set.
Solution Approach 2:
The patent introduces a latent space dimension that goes beyond the discrete patch dimension. By mapping faces to continuous vectors in latent space and performing operations in this higher-dimensional space, the system achieves continuous face generation at arbitrary angles and intermediate states.
2Reliability
If the number of learning images is increased to improve recognition rate, then recognition accuracy improves, but the requirement for more training data increases
Solution Approach 1:
The patent performs preliminary face generation during the training phase by creating synthetic faces at various angles and intermediate states. These pre-generated faces are added to the training set, so that when recognition is performed, the system already has diverse examples available without needing to collect additional real images.
Solution Approach 2:
The patent creates synthetic copies of faces by transforming existing training images through the learned model. These copied and transformed faces serve as additional learning images, effectively multiplying the available training data without requiring physical collection of more images.
3Object-affected harmful factors
If noise removal is performed on entire images, then overall noise reduction is achieved, but fine-detail noise in patches remains
Solution Approach 1:
The patent divides the image into patches and processes each patch separately through the denoising autoencoder. This segmentation allows the model to focus computational resources on local details and preserve fine structures that would be lost in global denoising operations.
Solution Approach 2:
The patent applies different processing strategies to different regions: entire image denoising for overall noise reduction and patch-level denoising for fine-detail preservation. The system adapts the denoising strength and method based on the local importance of the region, preserving edges and textures in critical areas.
Data Source
AI summary
At least some example embodiments disclose a device and a method for generating a synthetic image and a different-angled image and eliminating noise. The method may include receiving input images, extracting feature values corresponding to the input images using an image learning model, the image learning model permitting an input and an output to be identical and generating a synthetic image based on the feature values corresponding to the input images using the image learning model.