LiteDVDNet<T>

Models & Types · AiDotNet.Video.Denoising

LiteDVDNet high-speed video denoising (Ilchenko & Stirenko, IJIGSP 2025).

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

For Beginners

LiteDVDNet is a lightweight video denoiser designed for real-time performance. It achieves good denoising quality with significantly fewer parameters than full-scale models like DVDNet.

How It Works

References:

  • Paper: “LiteDVDNet: Optimizing FastDVDNet for High-Speed Video Denoising”, IJIGSP 2025
  • Builds on: Tassano, Delon & Veit, “FastDVDnet”, CVPR 2020

LiteDVDNet optimizes FastDVDnet for speed: it caches intermediate denoising results at inference, cuts the InputCvBlock’s intermediate channels from 90 to 30, simplifies each convolutional block to a single convolution, and halves the channel count (2.48M -> 0.64M parameters). LiteDVDNet-32 is 3x faster than FastDVDnet for -0.18 dB PSNR; LiteDVDNet-16 is 5x faster for -0.61 dB. It keeps FastDVDnet’s Conv -> BatchNorm -> ReLU ordering, PixelShuffle upsampling, and the residual connection between the central noisy input frame and the output. It does NOT use depthwise separable convolutions - earlier documentation here claimed it did, but that appears nowhere in the paper.

Example

// Create a lightweight LiteDVDNet model for efficient video denoising
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.ThreeDimensional,
    taskType: NeuralNetworkTaskType.Generative,
    inputHeight: 256, inputWidth: 256, inputDepth: 3);
var options = new LiteDVDNetOptions();
var liteDvdNet = new LiteDVDNet<double>(architecture, options);
            
// Or load a pre-trained ONNX model for real-time inference
var liteDvdNetOnnx = new LiteDVDNet<double>(architecture, "litedvdnet_model.onnx");

Constructors

Constructor Summary
LiteDVDNet(NeuralNetworkArchitecture<T> architecture, LiteDVDNetOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a LiteDVDNet model for native training and inference.
LiteDVDNet(NeuralNetworkArchitecture<T> architecture, string modelPath, LiteDVDNetOptions options) Creates a LiteDVDNet model for native training and inference.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
Tensor<T> Denoise(Tensor<T> noisyFrames)
Tensor<T> ForwardForTraining(Tensor<T> input) Applies the same residual on the tape-training path, so training optimizes the noise estimate that inference actually uses.