Skip to main content

Sheet layout and alignment

Normalized canvas​

alignImage first resizes the source image to width 1000 while preserving its aspect ratio. It converts the result to grayscale, applies a 5x5 Gaussian blur, applies adaptive Gaussian thresholding with block size 51 and constant 15, and performs a 5x5 rectangular morphological open.

It then searches external contours for four likely corner markers. A candidate must have:

  • Area greater than 100 and less than 3000.
  • Width-to-height aspect ratio between 0.7 and 1.3.
  • Contour extent greater than 0.6.

The four selected points are ordered by coordinate sums and differences, then mapped to destination points (50,50), (750,50), (750,1150), and (50,1150). The perspective warp produces an 800x1200 image.

Fallback​

If fewer than four candidate corners are found, the service logs a warning and resizes the original image directly to 800x1200. This fallback is not a configurable layout mode and does not establish accuracy for arbitrary sheet designs.

Fixed identity crops​

After alignment, OCR receives two fixed regions:

  • Name: (x=150, y=100, width=300, height=50).
  • Class: (x=500, y=100, width=150, height=50).

These coordinates assume the normalized 800x1200 canvas.