CNN based algorithm using MobilenetV2 for automated damage detection in concrete structures. Deployable on a mobile device, intended for drone surveillance images.
2. Overview
DRONE BASE
Provides professional drone
services to all industries
Aerial imaging and video
surveillances
Point of contact
Tomas Bucklin, Data Science Manager/
Engineer
Jason Rae, Full stack developer
William Murmann, Software Engineer
Goals
Optimized Machine learning
algorithm to detect structural
damage in real time.
Fast, efficient and Integrated on
the drone
6. Crack
detected
Undetected
Decks Cracked No crack
Pavements Cracked No crack
Walls Cracked No crack
TEST DATA
• Structural Defects Network
2018 (SDNET2018)
https://doi.org/10.15142/T3
TD19
• 56,000 images of cracked
and non-cracked concrete
bridge decks, walls, and
pavements
7. Methodology
Binary Image Classification
Damage detected Vs Undetected
If damage detected with threshold confidence,
assess further.
Convolutional Neural Network (CNN) via
Keras on top of TensorFlow
Demonstrated success in binary classifications
OR
8. Requirement
•Fast, efficient, deployable on cellphone
•Optimum battery usage (5-6 hours)
Damage
detection
•MobilenetV2: Designed & optimized for
a mobile device.
Convolutional
Neural
Network
(CNN)
10. Test Results
• Test accuracy ~ 80%
• Failed cases have very subtle
features
• Recommendation:
Multiclass Classification
CD
P
UD
Failed case
Success
12. Results
• Deliverable: Python module containing
model delivered
• Integration with Drone platform underway
• More data on Parking lot needed
Automated detections
Saves ∼2 weeks of human labor
3-4 days of official processing time.