PerceptualLoss<T>
Models & Types · AiDotNet.LossFunctions
Implements the Perceptual Loss function for comparing high-level features of images.
Inheritance: LossFunctionBase<T> → PerceptualLoss<T>
For Beginners
Perceptual Loss is a type of loss function used primarily in image processing and generative models. Unlike pixel-wise losses (like MSE) that compare images pixel by pixel, perceptual loss compares high-level features extracted from the images.
The key idea is to:
- Pass both the generated image and target image through a pre-trained network (like VGG)
- Extract features from various layers of this network
- Compare these features rather than raw pixels
This approach is more aligned with human perception because:
- It focuses on semantic content rather than exact pixel values
- It captures textures, patterns, and structures that are perceptually important
- It allows for some flexibility in pixel-level details while preserving overall appearance
Perceptual Loss is commonly used in:
- Style transfer algorithms
- Super-resolution models
- Image-to-image translation
- Any task where the “look” of an image is more important than exact pixel reproduction
Constructors
| Constructor | Summary |
|---|---|
PerceptualLoss(Func<Matrix<T>, Vector<Vector<T>>> featureExtractor, Vector<T> layerWeights) |
Initializes a new instance of the PerceptualLoss class. |
Methods
| Method | Summary |
|---|---|
T Calculate(Matrix<T> generated, Matrix<T> target) |
Calculates the Perceptual Loss between generated and target images. |
T CalculateLoss(Vector<T> predicted, Vector<T> actual) |
This method is not used for Perceptual Loss as it requires image matrices. |
Tensor<T> ComputeTapeLoss(Tensor<T> predicted, Tensor<T> target) |
|
void SetFeatureExtractorNetwork(INeuralNetwork<T> network) |
Sets the neural network used for tape-differentiable feature extraction. |