Modeling and Suppression of Inhomogeneous Intensity Edible Bird Nest Images for Impurity Segmentation Using β-Variational Autoencoder

This study proposes a β-variational autoencoder (β-VAE) method to address intensity inhomogeneity (IIH) in edible bird’s nest (EBN) images, which creates an uneven intensity that obscures fine impurity details and reduces segmentation accuracy. First, the β-VAE is used to learn the feature distribution of EBN images by mapping them into a latent space. This latent space is then disentangled through selective filtering and penalization of specific latent dimensions. This unsupervised learning approach effectively captures and isolates IIH in EBN images. Additionally, enabling precise segmentation of EBN and impurities without requiring annotated datasets. It also enhances robustness in handling unseen image instances. The proposed method achieves an intersection over union of 73.08% (equivalent to a Dice coefficient of 84.44%), surpassing existing segmentation techniques. By resolving IIH, this method improves the reliability and adaptability of automated EBN inspection systems for practical applications.