Applied Data Hackathon 2022.
Composite Vision team project on segmenting defects in CT scans of composite material parts.
When manufacturing Carbon fiber reinforced polymers, structural defects such as delaminations can occur inside material. These defects can impair composite physical properties and even render a detail unacceptably defective. Non-destructing testing is a procedure that can be used to identify these defects by means of computed tomography imaging, ultrasound imaging or other techniques.
In this project, we are working on a computer vision model capable of segmenting defects in CT images of a CFRP composite.
The data consists of CT images of several composite parts:
- 200 sample grayscale PNG images with annotations in VOC polygon format. Link to Yandex disk (24 Mb)
Image:
Annotations:
Segmentation map:
In this task, we have employed a modified U-Net architecture. The proposed architecture is illustrated below.
U-Net model:
For image labeling is used LabelMe.
For model prototyping -- Pytorch and torchvision.