Image Super-Resolution via Iterative Collaborative Filtering

The patent applies self-service by using the image's own low-resolution patches to generate high-resolution patches without external training data. The system extracts patches from the input image itself, processes them through the GAN-based super-resolution model, and uses the generated high-resolution patches to train the translation model, making the system self-sufficient and eliminating the need for large external datasets or prolonged training periods.

The patent extracts only the necessary components from the input image—specifically, low-resolution patches—that are required for super-resolution processing. Instead of storing or processing large external training datasets, the system extracts and processes only the relevant image patches from the input image itself, significantly reducing data storage requirements while maintaining enhancement quality.

The patent introduces a GAN-based translation model as an intermediary that learns the mapping from low-resolution to high-resolution patches. This intermediary model, trained on extracted image patches, enables the system to generate high-quality super-resolution images without extensively searching for matching patches, thus maintaining processing efficiency while significantly improving enhancement quality through learned transformations.

The patent segments the image into multiple patches and processes them independently through the super-resolution model. This segmentation approach reduces the computational burden on devices with limited resources, as each patch can be processed separately with smaller memory requirements, while still achieving high enhancement quality through the GAN-based translation model.