EXPLAINABLE MACHINE LEARNING-BASED COMPUTER VISION APPROACHES FOR ENGINEERING VISUAL CLASSIFICATION PROBLEMS

• 主要作者: Putra Suherman, Nayottama
• 格式: Theses
• 語言: Indonesia
• 主題: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• 在線閱讀: https://digilib.itb.ac.id/gdl/view/77383

With the development of computers and data availability, the data-driven approach of computer vision is emerging in popularity. However, the utilization of machine learning-based computer vision models has a black-box nature which is hard to justify their use in some sectors. With the still uncommon practice of qualitative evaluation, this thesis demonstrates the importance of feature-highlight investigation of the machine learning-based approach. This thesis compares three machine learning methods for image classification, which are principal component analysis, convolutional neural network, and capsule network, for solving three engineering cases. This thesis also highlights the insights gained from principal component analysis and investigates the compatibility of explainability methods with capsule networks. Out of the three techniques, it is found that the capsule network models are considered to be the best models in the three cases solved. It is also found that principal component analysis provides insight into the problem’s difficulty for all the data-driven approaches. By observing the results, it is found that the capsule network does not compatible with sensitivity analysis for explainability.