Gradient-Guided Super Resolution Using Residual LUT Kernels

The patent segments the super-resolution process into multiple stages: gradient calculation to determine direction types, directional kernel selection from predefined sets, and iterative refinement cycles. This segmentation transforms the complex deep learning task into manageable discrete operations that can be executed efficiently on mobile devices while maintaining image quality improvement.

The patent performs preliminary actions by pre-calculating and storing gradient direction types and selecting appropriate directional kernels before the actual super-resolution enhancement. By determining direction types based on gradient calculations in advance and preparing kernel sets accordingly, the system reduces real-time computational burden while preserving the ability to generate high-quality super-resolution images.

The patent changes parameters by using gradient-based direction types to select from multiple predefined kernel sets, rather than using a single fixed kernel or a large neural network. This parameter-driven approach allows the system to adapt to different image regions and directions while using a compact set of pre-computed kernels, significantly reducing memory requirements compared to storing large deep learning models on mobile devices.

The patent uses copying by storing pre-computed gradient direction information and kernel parameters in lookup tables rather than recalculating them during real-time processing. This copying of pre-prepared data structures enables fast retrieval and application during super-resolution enhancement, achieving real-time processing speeds while minimizing the memory footprint on mobile devices.