Video Frame Interpolation Model Using Cascaded Shared-Parameter Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional video frame interpolation methods often result in issues like ghosting, distortion, and blur, and deep learning-based methods can be prone to artifacts, while also consuming excessive computing resources and time.
Innovation Solution
A video frame interpolation method involving a cascaded training process using multiple reference models with shared parameters, where initial models are trained with specific loss functions and gradients to produce frame-interpolation results, optimizing the interpolation process without increasing network complexity or parameter count.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If traditional video frame interpolation methods are used, then the frame rate can be increased, but ghosting, distortion, and blur occur
Solution Approach 1:
The patent replaces traditional mechanical/optical interpolation methods with a deep learning-based neural network system. The video frame interpolation model uses learned features and patterns from training data to generate intermediate frames, substituting the mechanical frame blending approach with an intelligent system that can understand motion patterns and generate high-quality intermediate frames without ghosting or distortion.
Solution Approach 2:
The patent employs multiple loss functions (L1 loss, perceptual loss, gradient loss) during training to control different aspects of image quality. By adjusting the weights and parameters of these loss functions, the system optimizes the balance between frame rate enhancement and image quality preservation, reducing artifacts while maintaining temporal consistency.
2Manufacturing precision
If deep learning-based video frame interpolation methods are used, then image quality can be improved, but artifacts occur and computing resources are excessive
Solution Approach 1:
The patent implements a feedback mechanism through multiple loss functions that evaluate the generated intermediate frames against ground truth data during training. The L1 loss provides pixel-level feedback, perceptual loss provides semantic-level feedback, and gradient loss provides structural feedback. This multi-level feedback system guides the model to generate artifacts-free intermediate frames by continuously adjusting parameters based on comprehensive quality evaluation.
Solution Approach 2:
The patent combines multiple loss functions into a composite loss framework, where L1 loss, perceptual loss, and gradient loss work together synergistically. Each loss function addresses different types of artifacts and quality issues, and their combination creates a robust training objective that produces high-quality intermediate frames with minimal artifacts.
3Manufacturing precision
If deep learning-based video frame interpolation methods are used, then image quality can be improved, but computing time and resources increase
Solution Approach 1:
The patent segments the training process into distinct phases with different loss function combinations and weighting schemes. During different stages of training, different loss functions are emphasized to efficiently learn various aspects of frame interpolation. This segmented approach accelerates training convergence and reduces total computational time while maintaining high interpolation accuracy.
Solution Approach 2:
The patent applies different loss functions with varying weights depending on the training stage and specific requirements. Rather than applying all loss functions equally throughout training, the system selectively emphasizes certain loss components at different times, reducing unnecessary computational overhead while achieving optimal interpolation results.
4Manufacturing precision
If multiple reference models are used in training, then interpolation accuracy improves, but device complexity increases
Solution Approach 1:
The patent merges multiple reference models into a unified video frame interpolation model with shared parameters. Instead of maintaining separate models for different reference frames, the system integrates them into a single model that learns to process multiple reference inputs simultaneously. This merging reduces device complexity while maintaining the accuracy benefits of multi-reference processing through parameter sharing and unified architecture.
Data Source
AI summary
A video frame interpolation method and device, and a computer-readable storage medium are described. The method includes: inputting at least two image frames into a video frame interpolation model to obtain at least one frame-interpolation image frame, training the initial model using a first loss to obtain a reference model, copying the reference model to obtain three reference models with shared parameters, selecting different target sample images according to a preset rules to train the first/second reference model to obtain a first/second frame-interpolation result; selecting third target sample images from the first/second frame-interpolation result to train the third reference model to obtain the frame-interpolation result, obtaining a total loss of the first training model based on the frame-interpolation result and the sample images, adjusting parameters of the first training model based on the total loss, and using a parameter model via a predetermined number of iterations as the video frame interpolation model.