VideoDenoisingBase<T>

Base Classes · AiDotNet.Video

Base class for video denoising models that remove noise from video sequences.

Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → VideoDenoisingBase<T>

Implements: IShapeContract

Known derived types: BSVD<T>, FastDVDNet<T>, FloRNN<T>, LiteDVDNet<T>, ShiftNet<T>, UDVD<T>

For Beginners

Video denoising cleans up grainy or noisy video footage. This is common in low-light video, security camera footage, or video shot at high ISO settings. The model learns to distinguish real detail from random noise, removing the noise while keeping important details sharp. Using multiple frames helps because noise is random (different in each frame) while real content is consistent across frames.

How It Works

Video denoising removes noise and grain from video while preserving detail and temporal consistency. This base class provides:

  • Noise level (sigma) handling for both blind and non-blind denoising
  • Noise estimation utilities
  • Temporal buffer management for multi-frame denoising

Derived classes implement specific architectures like FastDVDNet, BSVD, FloRNN, etc.

Properties

Property Summary
bool IsBlindDenoising { get; set; } Gets whether this model performs blind denoising (estimates noise level automatically).
double NoiseSigma { get; set; } Gets or sets the noise sigma level for non-blind denoising.
int TemporalRadius { get; set; } Gets the temporal radius (number of frames before and after used for context).

Methods

Method Summary
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
LayerInputDomain GetInputDomain(int[] inputShape) Video denoisers accept public frame tensors on the unsigned 8-bit pixel scale.
LayerInputDomain GetOutputDomain(int[] outputShape) Denoised public outputs use the same unsigned 8-bit pixel scale as their inputs.
Tensor<T> Denoise(Tensor<T> noisyFrames) Denoises a sequence of video frames.
Tensor<T> ForwardForTraining(Tensor<T> input) Runs the native denoising stack on a public raw-frame tensor while keeping the tape objective in the normalized model domain.
double EstimateNoiseLevel(Tensor<T> noisyFrames) Estimates the noise level (sigma) from noisy input frames.
void Train(Tensor<T> input, Tensor<T> expectedOutput)