Image Enhancement via Region-Controlled Texture Synthesis

The patent segments the image enhancement process into distinct functional modules: a material image generating circuit that creates base texture patterns, and one or more texture generating circuits that apply these patterns to specific regions. This segmentation replaces the monolithic deep learning model with modular, computationally efficient components that can be independently controlled and optimized.

The patent applies different texture synthesis strategies to different regions of the image based on local characteristics. The system analyzes regional features such as edge density, texture complexity, and semantic content, then adapts the texture generation parameters accordingly. This local quality approach allows precise control over regenerated textures in different areas without requiring global deep learning processing.

The patent performs preliminary action by generating material images with appropriate texture characteristics before applying them to the compressed image. The material image generating circuit creates texture patterns with controlled frequency, orientation, and density properties in advance, allowing rapid application through simple synthesis operations rather than requiring complex post-processing of completely lost details.

The patent changes key parameters of the texture generation process including frequency spectrum characteristics, orientation angles, and spatial density to match the original image content. By adjusting these parameters in the texture generating circuits, the system can restore various types of details (edges, textures, patterns) without requiring computationally intensive deep learning inference.