BSVD<T>
Models & Types · AiDotNet.Video.Denoising
BSVD bidirectional streaming video denoising with real-time buffers.
Inheritance: NeuralNetworkBase<T> → VideoNeuralNetworkBase<T> → VideoDenoisingBase<T> → BSVD<T>
For Beginners
BSVD (Blind Spot Video Denoising) removes noise from video without needing clean reference frames. It uses a blind-spot network that learns denoising patterns directly from noisy video data.
How It Works
References:
- Paper: “BSVD: Bidirectional Streaming Video Denoising” (Qi et al., ACM MM 2022)
BSVD enables real-time video denoising through bidirectional streaming with efficient buffer management. It processes video in forward and backward passes, maintaining compact latent buffers for constant-memory operation regardless of video length.
Example
// Create a BSVD model for real-time bidirectional video denoising
var architecture = new NeuralNetworkArchitecture<double>(
inputType: InputType.ThreeDimensional,
taskType: NeuralNetworkTaskType.Generative,
inputHeight: 256, inputWidth: 256, inputDepth: 3);
var options = new BSVDOptions();
var bsvd = new BSVD<double>(architecture, options);
// Or load a pre-trained ONNX model for inference
var bsvdOnnx = new BSVD<double>(architecture, "bsvd_model.onnx");
Constructors
| Constructor | Summary |
|---|---|
BSVD(NeuralNetworkArchitecture<T> architecture, BSVDOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) |
Creates a BSVD model for native training and inference. |
BSVD(NeuralNetworkArchitecture<T> architecture, string modelPath, BSVDOptions options) |
Creates a BSVD model for native training and inference. |
Methods
| Method | Summary |
|---|---|
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input) |
Route the generic inspection path (used by GetNamedLayerActivations and test harnesses) through the same preprocessing that Denoise applies. |
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> Denoise(Tensor<T> noisyFrames) |
|
Tensor<T> ForwardForTraining(Tensor<T> input) |
Same rationale as GetNamedLayerActivations: the tape-based TrainWithTape path runs ForwardForTraining on the raw input. |