PSENetLoss<T>
Models & Types · AiDotNet.Document.OCR.TextDetection
PSENet’s training objective (Wang et al. 2019, Eq. 5): L = lambda L_c + (1 - lambda) L_s with lambda = 0.7, both terms Dice losses over the sigmoid of the kernel logits.
Inheritance: LossFunctionBase<T> → PSENetLoss<T>
How It Works
Layout is [batch, kernels, H, W] (or [kernels, H, W]), kernels ordered smallest to largest, so the LAST channel is the complete text map S_n. Dice is the paper’s D(S, G) = 2 sum(S G) / (sum S^2 + sum G^2), with the reference implementation’s 0.001 added to each denominator term, computed per image and per kernel.
L_c uses Online Hard Example Mining on the complete map: every positive pixel plus the 3x as many highest-scoring negatives. L_s scores the shrunk kernels only where the predicted complete map says text (S_n > 0.5), so they are not penalised outside a text region. The masks are constants; gradient flows through the scores alone, as in the reference.
Constructors
| Constructor | Summary |
|---|---|
PSENetLoss() |
Methods
| Method | Summary |
|---|---|
T CalculateLoss(Vector<T> predicted, Vector<T> actual) |
|
Tensor<T> ComputeTapeLoss(Tensor<T> predicted, Tensor<T> target) |
Fields
| Field | Summary |
|---|---|
double CompleteMapWeight |
Weight of the complete-text-map term (the paper’s lambda). |
int NegativeRatio |
Hard negatives kept per positive by OHEM. |