PADNet: Progressive-Difference-Aware Feature Reconstruction Mechanism for Anomaly Detection

Weiming Wang

Abstract: Unsupervised anomaly detection generally aims to identify irregularities using only normal samples, where feature reconstruction-based methods demonstrate greater robustness to noise by comparing reconstructed results with original data. However, they encounter issues with detailed information loss and insufficient anomaly discriminability. To address these challenges, we propose a progressive-difference-aware feature reconstruction network for image anomaly detection, named PADNet. To enhance context interaction, we develop a harmonic symmetric reconstruction framework integrated with a progressive feature harmonizer (PFH). The PFH mitigates detailed information loss to reduce undesired reconstruction errors through the progressive fusion of information flows. To enhance anomaly discriminability, we introduce the neighbor-aided residual feature representation module (NRFR) to strengthen difference-aware feature representations. The NRFR innovatively captures discriminative cues by interacting with neighboring reference samples in the feature cache pool. Experimental results on the MVTec, Visa, and BTAD datasets demonstrate that our method achieves superior performance while requiring only 25.3% of the parameters compared to the state-of-the-art baseline.