SQUAD: Scalar Quantized Representation Learning for Unsupervised Anomaly Detection and Localization

Abstract

The practical application of reconstruction-based models faces persistent challenges in enhancing reconstruction quality and accurately addressing anomalies, often called the “identical shortcut” phenomenon. This limitation has led to a gradual decline in the use of such methods. We introduce Scalar Quantized Unsupervised Anomaly Detection (SQUAD), an advanced framework based on the Vector Quantized Variational Autoencoder (VQ-VAE), to address these issues. SQUAD incorporates a novel quantization module, Finite Scalar Quantization (FSQ), alongside a discriminator, effectively overcoming problems like the identical shortcut phenomenon and codebook collapse. By leveraging advanced feature extraction and image reconstruction techniques, SQUAD achieves superior anomaly localization, consistently surpassing state-of-the-art methods across several benchmark datasets, including MVTecAD, ViSA, MPDD, BTAD, and KSDD2. Furthermore, SQUAD demonstrates superior results in image and pixel-level AUROC evaluations, mainly performing well in Pixel AP and PRO metrics.