Image Conversion Model Using Separation Outline and Shape Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for matching images of different types, whether using artificial intelligence algorithms or supervised learning techniques, face challenges such as low accuracy due to noise, differences in image style, and shape variations, as well as the need for human supervision and accurate training data.
Innovation Solution
A device and method for constructing a conversion model that transforms images of one type into another by using separate training data sets for each type, employing a generation module, separation outline module, shape inference module, and discriminator module to align and verify the converted images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If supervised learning algorithms (GAN) are used to convert images of different types, then image matching capability is improved, but human management and supervision are required and training data must be accurately matched
Solution Approach 1:
The system performs self-service by automatically training the conversion model using unaligned image pairs without requiring human supervision. The generator and discriminator modules work together in an automated adversarial training process, where the model learns to convert images between types and detect fake images independently, eliminating the need for human management and supervision while maintaining high matching capability
2Ease of manufacture
If training data with misaligned image pairs is used, then training process is simplified, but conversion accuracy decreases and noise is generated
Solution Approach 1:
The system converts the harmful effect of misaligned training data into a benefit by using the misalignment itself as the training objective. The adversarial training framework is specifically designed to handle unaligned image pairs, where the generator learns to create accurate conversions despite the lack of precise alignment, and the discriminator learns to detect inconsistencies. This approach transforms the previously problematic misaligned data into the ideal training material that drives the model to achieve high conversion accuracy while automatically filtering out noise and style variations
3Reliability
If multiple generators and discriminators are used to handle misaligned data, then conversion capability is improved, but device complexity increases
Solution Approach 1:
The system achieves universality by designing a single generator and a single discriminator that can handle all types of image conversion tasks and misalignment scenarios. The generator is trained to convert images from any first type to any second type, and the discriminator is trained to detect fake images regardless of the specific image types involved. This universal approach eliminates the need for multiple specialized generators and discriminators while maintaining comprehensive conversion capability across different image types and alignment conditions
Data Source
AI summary
A conversion model construction device updates a generation module to process a first training image of a first type into a conversion image by using a separation outline module, updates the generation module to process the conversion image into a shape of a second training image of a second type by using a shape inference module, and trains the conversion model by updating the generation module such that the conversion image is determined as a real image by a discriminator module. The separation outline module separates a polygon of an input image from a background and distinguishes an outline, the shape inference module compares a border shape of a polygon included in the first image with a border shape of a polygon included in the second image and the discriminator module determines whether the input image is a real image or a fake image according to the set condition.