Frame-Recurrent Video Super-Resolution for Temporal Consistency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video super-resolution methods are computationally expensive and produce temporally inconsistent results due to independent frame processing, leading to flickering artifacts.

Innovation Solution

A machine-learned recurrent super-resolution model using a frame-recurrent video super-resolution framework that leverages a previous HR estimate to super-resolve subsequent LR frames, reducing computational cost by warping only one image per step and ensuring temporal consistency through recurrent processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a batch of LR frames is combined to estimate a single HR frame using sliding window approach, then multi-frame information is utilized for super-resolution, but computational cost increases significantly and temporal consistency deteriorates

Engineering Contradiction:
Improvesuper-resolution qualityVSAvoidcomputational efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The video super-resolution task is segmented into independent frame-level estimations rather than processing large batches together. Each frame is estimated independently using a recurrent model that leverages temporal information from previous frames, avoiding the computational burden of processing entire batches while maintaining super-resolution quality.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The recurrent model maintains continuous temporal information across frames by incorporating previous HR estimates into current frame estimation. This continuous processing approach ensures temporal consistency while avoiding the discrete batch processing that causes computational inefficiency and flickering artifacts.

Inventive Principle:
Principle #20Continuity of useful action

2Device complexity

If each output frame is estimated independently conditioned on input frames, then frame processing is simplified, but temporal consistency deteriorates causing flickering artifacts

Engineering Contradiction:
Improveprocessing complexityVSAvoidtemporal consistency
Core Design Contradiction:
Device complexityVSStability of the object's composition

Solution Approach 1:

The model incorporates feedback from previously estimated HR frames into the current frame estimation process. By using the previous HR estimate as a condition for the current frame prediction, the model maintains temporal consistency and eliminates flickering artifacts while keeping the processing framework relatively simple.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The model performs preliminary estimation of HR frames in a sequential manner, where each frame estimation builds upon the previous one. This preliminary action of maintaining temporal context from prior frames ensures consistency throughout the video sequence without requiring complex post-processing.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3714424B1Frame-recurrent video super-resolution
Publication Date: 2025.10.08 GOOGLE LLC
  • EP3714424B1 patent drawingFigure 1A
  • EP3714424B1 patent drawingFigure 1B
  • EP3714424B1 patent drawingFigure 1C

AI summary

The present disclosure provides systems and methods to increase resolution of imagery. In one example embodiment, a computer-implemented method includes obtaining a current low-resolution image frame. The method includes obtaining a previous estimated high-resolution image frame, the previous estimated high-resolution frame being a high-resolution estimate of a previous low-resolution image frame. The method includes warping the previous estimated high-resolution image frame based on the current low-resolution image frame. The method includes inputting the warped previous estimated high-resolution image frame and the current low-resolution image frame into a machine-learned frame estimation model. The method includes receiving a current estimated high-resolution image frame as an output of the machine-learned frame estimation model, the current estimated high-resolution image frame being a high-resolution estimate of the current low-resolution image frame.