A Comparative Study of Transformer-Based Models and CNNs for Semantic Segmentation in Industrial CT Scans

Type: BA thesis

Status: running

Supervisors: Martin Leipert, Andreas Maier

Industrial computed tomography (iCT) is extensively utilized for non-destructive testing, material analysis, quality control, and metrology. In these applications, semantic segmentation is crucial, particularly for material analysis [1]. In recent years, Convolutional Neural Networks (CNNs) have been employed successfully for material segmentation, handling low-quality reconstructions, and performing complex segmentations where local context is vital. However, CNNs often struggle to capture long-range dependencies due to their localized nature.

Recently, transformer architectures have shown superior performance in various segmentation tasks. Unlike CNNs, which rely on filter banks to extract local features, transformers encode image patches into visual tokens and utilize self-attention mechanisms to process the entire input in parallel. This allows transformers to effectively capture long-distance relationships, though they may be less efficient when it comes to preserving fine local details [2]. In the context of three-dimensional (3D) data segmentation, both CNNs and transformers face challenges due to the increased memory requirements and the higher complexity of patterns that arise in 3D space.

One promising model for addressing these challenges is the Swin Transformer, which incorporates a hierarchical structure with shifted windows, enabling it to capture both local and global dependencies more efficiently [3]. To tackle the limitations of CNNs and transformers individually, hybrid models combining both architectures have been proposed. For instance, Cai et al. introduced a model that combines the Swin Transformer with CNNs for 3D segmentation tasks, taking advantage of each architecture’s strengths [2]. Both CNNs and Swin Transformers are known for their ability to generalize well on smaller datasets, thanks to their inductive bias [3] [4].

This thesis focuses on applying a hybrid approach combining the Swin Transformer and CNNs to a complex dataset of iCT scans of shoes, where the objective is to segment the shoes into their individual components, as demonstrated in previous work by Leipert et al. [5]. Despite the dataset’s small size, its high intrinsic variability and the relevance of both local and global dependencies make it an ideal candidate for evaluating segmentation methods. The segmentation will be performed in 3D, highlighting the challenges and opportunities of using these advanced models.

Through a comparative analysis of CNNs, the Swin Transformer, and their combined approaches, this thesis aims to provide insights into the strengths and limitations of each method in the context of 3D semantic segmentation on complex industrial CT datasets. The findings will contribute to improving segmentation techniques for iCT applications, potentially enhancing the accuracy and efficiency of material analysis in industrial contexts. The main limitation of the study is the application to a single dataset.

Literaturverzeichnis

[1] S. Bellens, P. Guerrero, P. Vandewalle and W. Dewulf, “Machine learning in industrial X-ray computed tomography – a review,” CIRP Journal of Manufacturing Science and Technology, vol. 51, pp. 324-341, 2024.
[2] Y. Cai, Y. Long, Z. Han, M. Liu, Y. Zheng, W. Yang and L. Chen, “Swin Unet3D: a three-dimensional medical image segmentation network combining vision transformer and convolution,” BMC Medical Informatics and Decision Making, vol. 23, p. 33, 2023.
[3] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin and B. Guo, “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,” in IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada, 2021.
[4] Y. Z. J. Z. D. Z. R. Y. Y. X. Xingwei He, “Swin Transformer Embedding UNet for Remote Sensing Image Semantic Segmentation,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-15, 2022.
[5] M. Leipert, G. Herl, J. Stebani, S. Zabler and A. K. Maier, “Three Step Volumetric Segmentation for Automated Shoe Fitting,” in 12th Conference on Industrial Computed Tomography (iCT) 2023, Fürth, Germany, 2023.