Defect Detection using SSD Network and Nvidia Jetson TX2 in Ultrasonic Inspection (Non-destructive Inspection)

Hello, I am Jason. I would like to share my project about the Defect Detection using SSD Network and Nvidia Jetson TX2 in Ultrasonic Inspection (Non-destructive Inspection).

This project aims to develop a system by using convolutional neutral network (CNN) to detect defects in the composite laminates automatically in order to increase the accuracy and efficiently in ultrasonic inspection.

1 Like

Ultrasonic testing (UT) is the one of most reliable methods to examine the serviceability of structural component, such as weld works, the composite light weight material and structure in aircraft and building structure.

However, ultrasonic testing is a manual task which is labor-intensive and time-consuming. The discontinuities and defects are usually no specific shapes, positions, and orientations. In fact, only skilled operators and qualified inspectors are certified to perform inspection and evaluation.

In order to achieve the strict standard of aircraft industrial and improve the current inspection procedure, a new ultrasonic inspection system with image analysis function using Single Shot MultiBox Detector (SSD) network and computer vision approach will be designed and developed to assist the inspector to improve the efficiency and accuracy for identifying the defect from interpreting the ultrasonic scanning image and reduce inspectors workload. A scanning system with objection detection function is applied on ultrasonic testing (UT) discontinuity indication.

1 Like

Identify discontinuities and classify the defects into two main categories by severity on the ultrasonic C-scan images Locate and size the discontinuities in the ultrasonic C-scan images Build a convolutional neural network-based defect detection model for ultrasonic C-scan Gather the network inference results and defect information in the inspection report


First, it needs to prepare the composite delamination defect samples. Then, it will perform the PAUT inspection on the prepared samples. Third, the C-scan images will be augmented of training dataset. Final, the images will be classified and labelled to be the dataset.


Official TensorFlow Docker image is downloaded and started by running the docker command. Verification is the important step to avoid system crash. bash shell & python session within a TensorFlow-configured container will be activated first. Then, TensorFlow will be imported. The Image of the Docker is Nvidia NGC Tensorflow 19.07-py3. At last, the TensorFlow program is ready to run and mount onto the directory.

After installing TensorFlow docker image from Nvidia NGC, the dependencies, such as the cython, contextlib2, matplotlib and user pandas are installed. Then, TensorFlow API, which the core of the training program, will be downloaded using the git clone command. Compile protobuf in research dir. Add PYTHONPATH in the directory of models/research and models/research/slim. Finally, validate the TensorFlow API successfully installed or not. If the TensorFlow API Installation is successfully installed, it is ready to train the customized model.