What is Crack Segmentation?

Crack segmentation is a computer vision task that labels each pixel of an image as crack or not crack. The output is a mask with the same size as the image. Crack pixels are set to 1 and all other pixels to 0.

How it differs from classification and detection

Classification gives one label for the whole image. Detection draws boxes. A box around a thin, winding crack is mostly background and says nothing about the crack’s width or branching. A mask keeps the shape of the crack.

What the mask gives

If the size of one pixel on the surface is known, the mask gives measurements:

• Area: the number of crack pixels multiplied by the area of one pixel.
• Length: the length of the one-pixel centreline of each crack, found by thinning the mask.
• Width: the distance from the centreline to the crack edge, doubled.

A coarser pixel size means that hairline cracks fall below one pixel and are lost. The pixel size on the surface depends on the camera and the distance to the pavement.

Models and data

Most crack models are encoder-decoder networks. The U-Net is the usual starting point. It has a contracting path that captures context and an expanding path that restores position. It trains from few images (U-Net paper ).

Public training sets are small and mostly show roads or campus pavement. DeepCrack has 537 annotated images of asphalt and concrete. CrackForest has 118 annotated road images, and CRACK500 has 500 phone images of pavement. CrackSeg9k combines several such sets. A model trained on them has not seen a grooved runway, so it needs checking on airfield images before its output is trusted.

What it does and does not do on an airfield

EASA guidance lists three pavement damage signs to look for during an inspection. They are cracking and spalling of concrete, the condition of joint sealing, and cracking of asphalt. It also notes that cracked and spalled pavement can shed pieces that become foreign object debris.

A mask is a measurement input, not a rating. In the PAVER method, the severity of a longitudinal or transverse crack in asphalt depends on spalling and filler condition. For a non-filled crack, a mean width above 6 mm also raises it to medium. Length is recorded in linear feet or metres per severity level.

Two airfield features need separate handling:

• Grooves: saw-cut grooves are transverse, 6 mm wide, 6 mm deep and 38 mm apart on a runway. A model has to tell these regular lines from cracks.
• Joints: concrete panels have joints at set spacing, between 3.8 m and 6.1 m depending on panel thickness. Joints are designed features and are not cracks, although the condition of the joint seal is inspected.