Airports runway defects detection

Solution for automatic defects detection (micro-size cracks and joints) on runways was developed

Deep learning
CUDA
Computer vision
Torch

2014

In cooperation with

About project

Task

Detect cracks and other hidden defects on an airfield surface.

Solution

  • Made software to build training set
  • Torch and CUDA-based model was developed
  • Web-dashboard to control GPU cluster

Small defects, big problems

Regular check of runway and surrounding area for cracks and other defects is an important part in airport life. Such problems should be identified at the first stage of their occurrence, as deeper destruction may lead to accidents, which is unacceptable in continuous passenger traffic.

At first glance it may seem that the task of finding defects on the runway is identical to the problem of finding defects on roads of Russia. The resemblance is definitely there: in the end, in both cases, the work is reduced to finding problems in the road surface. However, the airport task is complicated by several factors:

Regular road Airport runway
Surface Asphalt Concrete plates
Angle of shooting 30° 90°
Resolution of photo 1400 × 1400 px 16 000 × 16 000 px
Defects Holes, patches, cracks Microcracks
Size of a defect 5-200 cm < 5 mm
Importance Medium High
Table 1. Comparing the tasks finding defects on regular road and airport field

The table above shows that the size of defects we are interested in is much smaller than on regular roads. This creates additional problems not only in the work of preparing of the training sample, but also increases computational complexity.

130

times more
pixels

Pic 1. Comparing resolution of images

Model

To find really small defects we have developed a special configuration of a neural network with a high accuracy. Of course, increasing the machining accuracy increases execution time of the program and amount of noise. There are two tasks: to increase processing speed and remove extraneous noise from the results.

Pic 2. Runway surface before and after.
Pic 3. Runway surface before and after.
Pic 4. Close look

We was able to reduce execution time by several orders of magnitude due to the use of GPU NVidea through the use of technology CUDA, allowing us to transfer the calculation from CPU to GPU. This method allows us to parallelize calculations on thousands of concurrent threads.

Cores Performance
CPU 1-8
GPU (CUDA) 3000
Table 2. Comparing the tasks finding defects on regular road and airport field

Next project

3D tissue segmentation from CT