recognitionInput function

({int paddedWidth, Float32List tensor, int width}) recognitionInput(OcrImage crop, { int targetHeight = 48, int minWidth = 320, int maxWidth = 3200, })

Recognition preprocessing for one cropped text line: resize to the network's fixed targetHeight keeping aspect, then normalize to NCHW float32 in [-1, 1] (PP-OCR rec convention: (pixel/255 - 0.5) / 0.5).

The tensor is paddedWidth wide: the line's own width, but never less than minWidth (zero-padded on the right, like PaddleOCR's resize_norm_img, which pads to max(320/48, w/h) * 48). The recognizer is fully convolutional along x, so a long line keeps its aspect instead of being squeezed into a fixed width - squashing it is what turns legible characters into a different string. maxWidth is only a safety cap on pathological crops (the width it scales a wider line down to).

Returns the tensor, the width the line actually filled, and the tensor's width.

Implementation

({Float32List tensor, int width, int paddedWidth}) recognitionInput(
  OcrImage crop, {
  int targetHeight = 48,
  int minWidth = 320,
  int maxWidth = 3200,
}) {
  final scaledW =
      (crop.width * targetHeight / crop.height).round().clamp(1, maxWidth);
  final paddedW = scaledW > minWidth ? scaledW : minWidth;
  final resized = crop.resize(scaledW, targetHeight);
  final out = Float32List(3 * targetHeight * paddedW); // zero-padded
  final plane = targetHeight * paddedW;
  final rgba = resized.rgba;
  for (var y = 0; y < targetHeight; y++) {
    for (var x = 0; x < scaledW; x++) {
      final p = (y * scaledW + x) * 4;
      final o = y * paddedW + x;
      out[o] = (rgba[p] / 255.0 - 0.5) / 0.5;
      out[plane + o] = (rgba[p + 1] / 255.0 - 0.5) / 0.5;
      out[2 * plane + o] = (rgba[p + 2] / 255.0 - 0.5) / 0.5;
    }
  }
  return (tensor: out, width: scaledW, paddedWidth: paddedW);
}