Cascaded Back Projection Networks for Data-Faithful Super-Resolution Images

The patent divides the super-resolution process into multiple stages using a cascaded network architecture. Instead of a single fixed up-sampling operation, the image is processed through multiple sequential stages (first network, second network, etc.), each performing a portion of the up-sampling task. This segmentation allows each stage to focus on specific frequency components or resolution levels, preventing the blurring that occurs in single-stage fixed up-sampling while maintaining data fidelity.

The patent employs dynamic residual learning where each network stage adapts to the specific characteristics of the input image and previous outputs. The residual connection mechanism allows the network to dynamically adjust its behavior based on the input, learning optimal up-sampling transformations rather than applying fixed pre-determined transformations. This dynamic approach prevents blurring by adapting to the actual image content and maintaining fidelity to the original low-resolution input.

The patent segments the super-resolution task into multiple stages, with each stage handling a specific portion of the up-sampling process. The first network performs initial up-sampling to an intermediate resolution, and subsequent networks continue the process to achieve the final high resolution. This segmentation allows each stage to be computationally efficient while collectively achieving high image quality that would be difficult to obtain in a single stage.

The patent implements continuous refinement through residual connections, where each network stage builds upon the previous stage's output. The residual learning mechanism ensures that each stage contributes useful information to the final output, and the continuous nature of this refinement process allows the system to achieve high image quality progressively rather than requiring a single complex stage.

The patent applies preliminary super-resolution processing to transform low-resolution captured images into high-resolution images before storage and transmission. By performing this up-sampling operation in advance using the cascaded network, the system reduces the amount of data that needs to be stored and transmitted while maintaining high data fidelity. The residual learning mechanism ensures that the transformation preserves essential image information.

The patent changes the resolution parameter of the image data through the cascaded network processing. The system transforms images from low resolution to high resolution by applying multiple up-sampling operations in sequence, each building upon the previous one. This parameter change allows the system to maintain data fidelity equivalent to capturing high-resolution images directly, while significantly reducing the actual data volume that needs to be stored and transmitted.