Normality Addition via Normality Detection in Industrial Image Anomaly Detection Models

TL;DR

This work tackles changing normality criteria in industrial image anomaly detection by introducing Normality Addition via Normality Detection (NAND), a post-hoc, text-guided adaptation powered by vision-language models. NAND detects the newly defined normality from textual descriptions, uses a Prompt Generator to create meaningful prompts, and leverages a Normality Detection Module based on APRIL-GAN to generate a suppression map that reshapes the anomaly map as $A_{final} = A \odot (1 - A_{sup})$, yielding a revised anomaly score $a = \max A_{final}$. The method is model-agnostic and demonstrated on MVTec AD with multiple detectors (APRIL-GAN, WinCLIP, PatchCore), showing consistent AUROC gains (e.g., $\text{AUROC}_{\text{before}} = 70.6$ to $\text{AUROC}_{\text{after}} = 75.1$ for APRIL-GAN) and selective suppression of the added normality while preserving other anomalies. Overall, NAND provides a practical pathway for rapid, text-guided adaptation of industrial IAD systems to evolving definitions of normality without retraining.

Abstract

The task of image anomaly detection (IAD) aims to identify deviations from normality in image data. These anomalies are patterns that deviate significantly from what the IAD model has learned from the data during training. However, in real-world scenarios, the criteria for what constitutes normality often change, necessitating the reclassification of previously anomalous instances as normal. To address this challenge, we propose a new scenario termed "normality addition," involving the post-training adjustment of decision boundaries to incorporate new normalities. To address this challenge, we propose a method called Normality Addition via Normality Detection (NAND), leveraging a vision-language model. NAND performs normality detection which detect patterns related to the intended normality within images based on textual descriptions. We then modify the results of a pre-trained IAD model to implement this normality addition. Using the benchmark dataset in IAD, MVTec AD, we establish an evaluation protocol for the normality addition task and empirically demonstrate the effectiveness of the NAND method.

Figures (4)

• Figure 1: A schematic description of (a) CLIP clip2021icml and (b) APRIL-GAN aprilgan.
• Figure 2: A depiction of the proposed method, NAND. For a given anomaly detection model ($\mathcal{A}_\theta$) and a text-based normality, NAND starts by generating text prompts for the inputs to the ND module. The output of ND module, a suppression map, is element-wise multiplied to suppress the output of the given anomaly detection model.
• Figure 3: Examples of similar anomaly types in MVTec AD mvtecad dataset.
• Figure 4: Anomaly maps for input images before and after applying NAND to the anomaly detection model.