Publication: Modeling and suppression of inhomogeneous intensity in edible bird nest images for impurity segmentation using β-variational autoencoder
Date
2024-12-01
Authors
Khairul Firdaus, Mohd Talib
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Edible bird’s nest (EBN) exhibits complex optical properties that thrived the
occurrence of intensity inhomogeneity (IIH) in its digital images. This hinders the
performance of automated EBN visual inspection which is still under development.
Previous studies addressed this issue via several methods including image fusion,
lighting manipulation study, and recently employed deep learning frameworks. While
deep learning offers advantages over conventional methods, existing studies
predominantly focused on impurity detection. The implementations are constrained by
the needs of a well-crafted annotated datasets. The modeling of EBN IIH distribution
remains unexplored, where specific features of EBN and impurities can be modeled
and suppressed for segmentation. This study utilized a deep generative model, β-
Variational Autoencoder (β-VAE), to learn semantic image features and map them into
distinctive distributions in the latent space. This space was then analysed and
disentangled with a latent disentanglement method proposed in this work, which
involves selective filtering and penalizing of latent dimensions. The method addressed
the trade-off when using a small β value which preserves the latent interpretability in
VAE. A more refined image is attained during image reconstruction with meaningful
features retained for better segmentation. An overall IoU score of 73.08% (equivalent
to a Dice coefficient of 84.44%) was achieved, surpassing the performance of existing
studies recorded at 74.61%. The study has contributed to the modeling of IIH distribution in EBN for impurities segmentation. It also mitigated the need of
annotated datasets which has been a limitation in previous study.