Guided CNN Frame Restoration for Low-Complexity Video Coding

The patent introduces an intermediary constrained subspace that mediates between the complex restoration task and the limited model capacity. By projecting the restoration problem into a predefined subspace spanned by basis images, the system achieves accurate restoration without requiring complex neural network architectures. The subspace acts as an intermediary representation that simplifies the restoration process while maintaining high accuracy.

The patent transforms the restoration problem from learning complex pixel-wise transformations to learning parameters (coefficients) in a predefined subspace. Instead of directly predicting restored pixels, the model learns to predict coefficients that combine basis images to form the restored output. This parameter transformation simplifies the learning task and reduces model complexity while preserving restoration accuracy.

The constrained subspace serves as an intermediary that reduces computational resources required for restoration. By limiting the solution space to a predefined subspace with fixed basis images, the system avoids the need for complex, resource-intensive neural network architectures while maintaining high restoration quality. The intermediary subspace enables efficient computation through simple coefficient combination operations.

The patent replaces expensive, complex neural network models with a simpler, more resource-efficient approach using predefined basis images and coefficient prediction. The complex model is substituted with a lightweight system that combines simple basis images through learned coefficients, achieving comparable restoration quality with significantly reduced computational resources and energy consumption.

The patent converts the harmful artifacts introduced by lossy quantization into a beneficial restoration opportunity. Instead of simply accepting the degraded reconstructed image, the system uses the artifacts as input signals that the neural network learns to correct. The constrained subspace restoration process transforms the quantization damage into an opportunity for intelligent recovery, converting harm into benefit by leveraging the artifacts as cues for restoration.

The patent implements a feedback mechanism where the reconstructed image (containing quantization artifacts) is fed into the neural network, which then generates a restored version. The network learns from the difference between the degraded input and the original clean images, creating a feedback loop that continuously improves restoration quality. This feedback process enables the system to compensate for quantization losses and recover image quality while maintaining compression efficiency.