Quality Assessment of Production Using Image Segmentation

Abstract

Quality assessment of manufactured products is vital to ensure performance, safety, and customer satisfaction across industries. Defects in items such as bottle caps, cables, capsules, leather, and metal components can affect functionality and durability. Traditional inspection methods relying on manual visual checks are time-consuming and error-prone. This study proposes an AI driven framework using the Probabilistic U-Net integrated with a Conditional Variational Autoencoder (CVAE) for automated defect detection. The model introduces stochastic latent variables to generate multiple plausible segmentation maps, enhancing accuracy under ambiguous or noisy conditions. Using the MVTec Anomaly Detection dataset, which includes defects such as scratches and discoloration, the system applies preprocessing steps including resizing, normalization, and data augmentation to enhance the robustness and consistency of the input data. A hybrid loss combining cross-entropy and Kullback–Leibler divergence improves segmentation precision and latent space alignment. Experimental results confirm robust and reliable defect detection across diverse product categories, demonstrating the model’s potential for automated manufacturing quality assurance.