Image Segmentation of the γ'-Phase in Nickelbase Superalloys utilising Deep Learning

Image Segmentation of the γ'-Phase in Nickelbase Superalloys utilising Deep Learning

Riegger, Franzi (TU Delft Electrical Engineering, Mathematics and Computer Science)

Möller, Matthias (mentor)
von Lautz, Julian (mentor)

Delft University of Technology

2019-09-10

Quantitative analysis of material microstructure is a well-known method to derive chemical and physical properties of a sample. This includes the segmentation of e.g. Light Optical Microscopy or Scanning Electron Microscopy images where each pixel is assigned to a material. Since some phases such as the γ-γ’ structure in nickelbased superalloys feature a high complexity and appear in a wide range of different sizes and shapes, segmentation is traditionally done by experts in a manual ap- proach . With Deep Neural Networks, a competitive method to automate this time-consuming, costly and inconsistent analysis, is found. Recently, their special instances of Convolutional Neu- ral Networks outperform any other machine learning algorithms in computer vision tasks by com- pressing the raw input image content to its most meaningful features. Applying these models to image segmentation demands for an adaption of their architecture, in order to maximize both, lo- cation and semantic information. Current research mainly focuses on two possible modifications, namely extending the encoding Convolutional Neural Network by a decoding substrcuture or partly substituting it by a context module. Maintaining the structural subdivision of encoder, context and decoder module allows a flexible, construction-kit like definition of segmentation models. The Ten- sorFlow library offers a framework for implementation. Based hereon, novel architectures are inves- tigated. Moreover, an insight into the impact of the single subelements and particular combinations of these is gained. For this, a chosen subsets of possible models are trained on 206 images. Com- prehensive understanding of internal mechanisms is achieved by visualization, error analysis and generalized consideration of segmentation relevant model characteristics such as the Effective Re- ceptive Field. The model selection is based on the validation results, derived from segmenting fur- ther 84 images. On this data set, the final model predicts the ground truth data with almost human like accuracy.

Convolutional Neural Networks
Image Segmentation
Deep Learning

http://resolver.tudelft.nl/uuid:62366de2-df60-4028-b541-9485cc6b7e2c

Student theses

master thesis

© 2019 Franzi Riegger