No-MambAAD : Revitalizing Conv-Only Networks for Unsupervised Anomaly Detection

Abstract

Most of the current state-of-the-art visual unsupervised anomaly detection (UAD) methods leverage complex neural architecture modules: Transformer-based methods provide high-quality anomaly detection performance due to their global feature extraction capability, similar to the recent Mamba based methods that combine the strengths of CNNs and Transformers. Some of the simpler reconstruction-based UAD methods are purely CNN-based, which offers linear complexity, but is performance-restricted by feature extraction locality. Hence, the architecture variants have inherent design trade-offs: CNNs lacks long-range feature interaction, Transformers struggle with quadratic complexity, and Mamba based solutions suffer in high parameter count and scalability. In this work we propose to revisit CNN-based approaches by introducing novel stripmodulation and gated-mixer mechanisms, and propose No-MambAAD, a novel visual UAD method absent of Mamba and Attention blocks. The proposed method offers similar or better anomaly detection performance than the current state-of-the-art approaches and outperforms the current state-of-the-art across multiple benchmarks with 38% smaller parameter count.