EDVR<T>

Models & Types · AiDotNet.Video.Enhancement

EDVR: Video Restoration with Enhanced Deformable Convolutional Networks.

Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → VideoSuperResolutionBase<T> → EDVR<T>

For Beginners

EDVR is a state-of-the-art video restoration model that can:

  • Upscale low-resolution video (super-resolution)
  • Remove blur and noise (deblurring/denoising)
  • Handle complex motions and occlusions

The model uses alignment to compensate for motion between frames, then fuses information from multiple frames to restore high-quality output.

Example usage:

How It Works

Technical Details:

  • PCD (Pyramid, Cascading and Deformable) alignment module
  • TSA (Temporal and Spatial Attention) fusion module
  • Deformable convolutions for motion-adaptive feature alignment

Reference: “EDVR: Video Restoration with Enhanced Deformable Convolutional Networks” CVPR 2019 https://arxiv.org/abs/1905.02716

Constructors

Constructor Summary
EDVR() Initializes a new instance with default architecture settings.
EDVR(NeuralNetworkArchitecture<T> architecture, EDVROptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction)
EDVR(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, EDVROptions options)

Properties

Property Summary
bool SupportsTraining { get; }

Methods

Method Summary
List<Tensor<T>> Enhance(List<Tensor<T>> frames) Enhances video frames with super-resolution and restoration.
ModelMetadata<T> GetModelMetadata()
Tensor<T> EnhanceFrame(List<Tensor<T>> neighborFrames) Enhances a single frame using neighboring frames.
Tensor<T> Upscale(Tensor<T> lowResFrames)