The Technology to Detect Corrosion in Bridge-Attached EquipmentsNTT Access Network Service Systems Laboratories

This technology is automatically detects corrosion from images of bridge-attached equipments.
This technology can improve the workability of deterioration diagnosis because it can automatically detect the corrosion that has occurred in the Bridge-attached equipment by image processing using photographed image such as taken UAV and telephoto cameras.

NTT Group has bridge-attached equipments on about 43 thousand bridges in japan and periodically inspects all of those.
In current inspection and diagnosis, the operator takes images of the equipment to be inspected and manually checks all images for deterioration and marks the deteriorated areas.
The current method is challenged by individual differences in operators and the operational aspects of checking the deterioration of a large amount of equipment, so there is a need for more efficient maintenance.

• The combination of two algorithms (area recognition and automatic detection of corroded areas) provides corrosion detection with only area recognition of the area to be inspected.

• Used as a tool at the time of judging faults in the work for inspecting and diagnosing bridges

When detecting corroded areas from images, there was a problem of over detection of areas other than those to be inspected.
With this considered, Algorithms for detecting corrosion areas from images (1) and for identifying the equipment to be inspected (2) were developed, and, by taking the logical conjunction of (1) and (2), it is possible to accurately detect the corrosion of the equipment.
Items (1) and (2) are constructed in the following method.

(1) The detection of corroded areas uses an image segmentation method to build a corrosion area detection model by learning to detect corrosion pixel by pixel from the images. Iterative learning in the construction of a detector in the corrosion region implements AUC maximization to suppress over-detection of corrosion.

(2) The detection of Equipment areas uses a CNN-based image classification method to build an equipment area detection model by learning to detect inspection equipments from the images. The equipment detection accuracy is enhanced by optimizing the images sizes for the recognition of equipment areas.

bridge attached equipment
This is the equipment for the conduits housing cables to cross a river or the like, and, comprises with an conduits and the attached members to support it.

CNN
Convolutional Neural Network

Image segmentation
This is one of the image recognizing technologies in deep learning. Performing pixel-by-pixel recognition from images.

AUC
Area Under a Receiver Operating Characteristic Curve

NTT Access Network Service Systems Laboratories - Civil Engineering Project