DBNet<T>

Models & Types · AiDotNet.ComputerVision.Detection.TextDetection

DBNet (Differentiable Binarization Network) text detector.

Inheritance: ModelBase<T, Tensor<T>, Tensor<T>> → VisionTaskModelBase<T> → TextDetectorBase<T> → DBNet<T>

For Beginners

DBNet is a state-of-the-art text detector that uses differentiable binarization to segment text regions. Unlike traditional methods that use a fixed threshold, DBNet learns an adaptive threshold for each pixel, making it more robust to varying text appearances.

How It Works

Key features:

  • Differentiable binarization for end-to-end training
  • Adaptive thresholding per pixel
  • Fast inference with single-pass architecture
  • Works well for both regular and irregular text shapes

Architecture, as in the paper and its reference implementation: a ResNet backbone; a feature pyramid whose four levels are reduced to a common width by 1x1 lateral convolutions, merged top-down, smoothed by 3x3 convolutions to a quarter of that width and upsampled to 1/4 resolution and concatenated; then two identical heads - convolution, batch norm, ReLU, then two stride-2 transposed convolutions back to full resolution - predicting the probability map and the threshold map.

Reference: Liao et al., “Real-time Scene Text Detection with Differentiable Binarization”, AAAI 2020

Constructors

Constructor Summary
DBNet(TextDetectionOptions<T> options, double k) Creates a new DBNet text detector.

Properties

Property Summary
string Name { get; }

Methods

Method Summary
Task LoadWeightsAsync(string pathOrUrl, CancellationToken cancellationToken)
void SaveWeights(string path)
void SetTrainingMode(bool training)