Image Crop Enhancement via GAN Resolution Scaling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image databases suffer from heterogeneous image quality due to resolution limitations of collection devices and compression algorithms, resulting in limited thematic resources and versatility.
Innovation Solution
A computer-implemented method that selects image portions, identifies similar images, determines enhancement scores, and increases pixel resolution using a Generative Adversarial Network (GAN) tool to create enhanced image crops, which are stored in the database when the synthetic value is below a tolerance value, thereby improving image quality and resource diversity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image resolution is increased using traditional interpolation methods, then pixel resolution is improved, but image quality and realism deteriorate due to artificial appearance
Solution Approach 1:
A Generative Adversarial Network (GAN) is introduced as an intermediary system between the low-resolution input image and the final high-resolution output. The GAN consists of two neural networks (generator and discriminator) that work together to produce realistic enhanced images, bridging the gap between simple interpolation and complex image reconstruction
Solution Approach 2:
The patent transforms the image enhancement problem by changing parameters from traditional interpolation coefficients to neural network weights and biases. The system learns optimal parameter transformations through training on paired low-resolution and high-resolution images, enabling adaptive resolution enhancement that preserves image quality
2Reliability
If all images in the database are enhanced to high resolution, then image quality is improved, but computational resources and processing time increase significantly
Solution Approach 1:
The patent applies resolution enhancement selectively rather than uniformly across all database images. The system identifies specific images or regions that benefit most from enhancement, applying the computationally intensive GAN process only where necessary, while leaving other images in their original form
Solution Approach 2:
The system performs partial enhancement by focusing computational resources on enhancing only certain portions of the database or specific critical regions within images, rather than processing every pixel of every image at full resolution, thus reducing overall computational burden
3Quantity of substance
If the database stores only original low-resolution images, then storage space is conserved, but thematic resources and versatility are limited
Solution Approach 1:
The patent segments the image database into different resolution tiers and quality levels. Original low-resolution images are retained for storage efficiency, while selectively enhanced high-resolution versions are created and stored for specific applications, allowing the system to serve different quality requirements from the same database infrastructure
Solution Approach 2:
The system creates enhanced copies of selected images rather than replacing originals. The GAN-generated high-resolution versions are stored as separate entities in the database, allowing users to access either the original compact version or the enhanced version depending on their needs, effectively multiplying the utility of each stored image
Data Source
AI summary
A method including selecting, in a server, a first image portion from an image is provided. The method also includes identifying one or more known similar images associated with the first image portion, and determining a first score for enhancing the first image portion based on the known similar image(s). The method includes increasing a pixel resolution in the first image portion according to the scale to form an enhanced image portion. The method also includes identifying a synthetic value for the enhanced image portion and storing the enhanced image portion in a database when the synthetic value is below a tolerance value.