Explainable Visual Anomaly Detection via Concept Bottleneck Models

This repository contains the code for our work "Explainable Visual Anomaly Detection via Concept Bottleneck Models", through which we try to enhance interpretability in Industrial Visual Anomaly Detection (VAD) by employing Concept Bottleneck Models (CBMs), i.e. neural models that learn meaningful concepts and thus can provide human-understandable descriptions of anomalies and defects. Together with training and testing the standard CBM architectures, we propose a pipeline for generating synthetic anomalies that enables effective concept learning without requiring anomalous samples, thus maintaining traditional VAD requirements. Finally, we propose a modification to the classic CBM architecture to produce not only concept-based explanations but also visual explanations, similar to those in standard VAD models, by integrating a novelty detection branch that leverages features-based VAD algorithms. The experiments were performed starting from the MVTec-AD dataset, which was modified to integrate generated images and concept annotations.

How to use

The repository is organized as follows:

• /datasets defines the classes for the original MVTec dataset and the final concept dataset that was used for training the CBMs;
• /models contains the scripts necessary to build the models;
• /trainers and /evaluators contain the scripts for respectively training and evaluating both the CBM models and the novelty detection branch for the generation of a heatmap;
• /main_scripts contains the scripts for performing data annotation with concepts, CBM training/testing, novelty detection training/testing, concept evaluation and intervention over concepts.

See main.sh for an example of how to run the main scripts.

Dataset

We tested our dataset creation pipeline starting from the MVTec-AD dataset, which was modified to include artificially generated anomalous images, together with the original ones, and concept annotation. To replicate our pipeline for concept extraction and annotation, the MVTec-AD dataset should be stored according to the following structure:

├── mvtec
│   ├── bottle
│   ├── cable
│   │   ├── ground_truth
│   │   ├── test
│   │   │   ├── bent_wire
│   │   │   │    ├── 000.png
│   │   │   ├── generated anomalies
│   │   │       ├── bent_wire
│   │   │           ├── 000.png
│   │   ├── train
│   │   │   ├── good
│   │   │       ├── 000.png
│   ├── capsule
│   ├── ...

Paper Abstract

In recent years, Visual Anomaly Detection (VAD) has gained significant attention due to its ability to identify anomalous images using only normal images during training. Many VAD models work without pixel-level supervision, but are still able to provide visual explanations by highlighting the anomalous regions within an image. However, though these visual explanations can be helpful, they lack a direct and semantically meaningful interpretation for users. To address this limitation, we propose adapting Concept Bottleneck Models (CBMs) to the VAD setting. By learning meaningful concepts, the network can provide human-interpretable descriptions of anomalies, offering a novel and more insightful way to explain them. Our contributions are threefold: (i) we develop a Concept Dataset to support research on CBMs for VAD; (ii) we modify the CBM architecture to generate both concept-based and visual explanations, bridging interpretability and localization; and (iii) we introduce a pipeline for synthesizing artificial anomalies, preserving the VAD paradigm of training solely on normal data. Our approach, Concept-Aware Visual Anomaly Detection (\ourmethod{}), achieves performance comparable to classic VAD methods while providing richer, concept-driven explanations that enhance interpretability and trust in VAD systems.