Machine and Deep learning to detect cracks on the road surface from 20M pixel images and report them geospacially.
The defects to be detected are dependent on the road type. There are three basic road types:- 1. Bituminous, 2. Concrete and 3. Bituminous overlaid concrete.
CRACKING
All road types can have cracking. This can be longitudinal (along the road) or transverse (across the road) or random. The random cracking can form crazed areas with many cracks.
The software needs to detect all forms of cracking and be able to show the detected ones as an overlay on the image. Images should be stored - overlaid image and cracks only image. The full image should not be analysed. An investigation area should be specified. The next stage is reporting and quantifying the results. Transverse cracks are the simplest and chainage in metres and coordinates(centre position) need recorded. Longitudinal cracks need a start and end chainage and their cross-sectional position recorded. The simplest way to record cross-sectional is make five zones in the investigation area working from the top to the bottom of the image and called A, B, C, D and E so and example of a longitudinal crack may be Longitudinal Crack 15 to 45m in zone B. Random cracks are more difficult to deal with. The amount of cracking is the significant factor and I propose three levels - minor, major and severe. The area of the cracking can be specified by chainage and cross sectional zones.
I have photos and ground truth dataset, only i need to extract user preference and POI properties as visual content using CNN.