Unsupervised Detection of Visual Anomalies in Industrial Images Using Convolutional Autoencoders

Automatic visual inspection is an essential component of industrial manufacturing systems, where even small surface defects can lead to significant financial and reputational losses. This paper presents an unsupervised approach for detecting visual anomalies using convolutional autoencoders (CAEs) trained exclusively on defect-free industrial images. Three models of increasing depth were designed and trained on 269 grayscale samples at a resolution of 512×512. Reconstruction-based error metrics were used to quantify the difference between input and output images, including mean squared error (MSE), structural similarity (SSIM), mean absolute error (MAE) with Gaussian pre-blur, and MSE with Gaussian pre-blur. The decision threshold was estimated from the validation distribution of normal data using the 95th percentile criterion. The results indicate that shallow networks achieve higher precision but fail to detect a large portion of anomalies, whereas deeper networks tend to overgeneralize and reconstruct defective regions. Although none of the models reached production-level reliability, the study highlights the main trade-offs in architecture design and provides a reproducible baseline for future anomaly detection research in industrial image inspection.