SAMPE Journal - May/June 2020 - 37

said network. This allows for any potential application that wishes to perform
analysis from the defect data to be both theoretically possible and easily constructed on top of the infrastructure provided. This database is not an altered
version of the IMT database, but rather a standalone one.
Image Analysis
Machine Learning Approach
The ML approach described in this paper is based on the network architecture
outlined in13. The traditional method of object detection, the classification of
patches of an image, simply does not have high enough fidelity for the post-inspection applications outline previously. Thus, rather than patch classification, our approach attempts to assign each individual pixel a defect classification. This is accomplished by replacing the standard CNN architecture with all
convolutional layer, creating what is known as a Fully Convolutional Network
(FCN). Thus, the ability to have an accurate and detailed representation of a
given defect is limited only by the resolution of the scanning or imaging system
utilized in the data acquisition system.
Looking to further improve the potential classification accuracy of the system, a FCN variant of the ResNet neural network architecture4 was constructed. ResNet is notable for having scored a 3.6% top five error rate for the ILSVRC
image recognition competition. The network has a total of 15 of the ResNet
"skip functions", with 3 convolutional layers allocated to each skip function,
bringing the total network size to 45 convolutional layers. Each skip function
takes the output of a layer and adds it to the input of a layer further on in the
network. This allows for a network to be built that consists of many layers, but
this have a comparatively low parameter count. Glorot initialization14 and convolutional batch normalization were used due to their effectiveness in improving the performance of convolutional networks.
The training was accomplished using an Nvidia Titan Xp GPU due to the
ability of a capable GPU to rapidly accelerate the training of deep network
models such as ours15,16. Approximately 500 800x800 pixel scan images were
used for the training dataset, with testing and validation of datasets of 10 each.
In addition, the use of a live system has been tentatively examined and preliminary results of such tests are positive.

De-Comp
Composites, Inc.

Composite Materials
& Supplies

A One-Stop
Shopping Center
Carbon, Fiber Glass,
& Kevlar Fabrics
Bagging Film
Breather/Bleeder
Release Films
Release Fabrics/Peel Plies
Tapes
Vacuum Valves & Hoses
Sealant Tapes
Release Liquids
Cutting Tools

Figure 3. UI Startup window showing the functionalities.

Resins
Tooling Fabrics, Structures,
& Supplies
Many, Many More Products
E-mail:
email@decomp.com
Website:
www.decomp.com
1519 Eastgate Drive
Cleveland, OK 74020
P: +1 918.358.5881 | F: +1 918.358.3750
w w w. s a m p e . o r g

M AY/J U N E 2 0 2 0

SAMPE JOURNAL

http://www.decomp.com http://www.sampe.org

SAMPE Journal - May/June 2020

Table of Contents for the Digital Edition of SAMPE Journal - May/June 2020