Learning Feature Inversion for Multi-class Anomaly Detection under General-purpose COCO-AD Benchmark (2404.10760v1)

Abstract: Anomaly detection (AD) is often focused on detecting anomaly areas for industrial quality inspection and medical lesion examination. However, due to the specific scenario targets, the data scale for AD is relatively small, and evaluation metrics are still deficient compared to classic vision tasks, such as object detection and semantic segmentation. To fill these gaps, this work first constructs a large-scale and general-purpose COCO-AD dataset by extending COCO to the AD field. This enables fair evaluation and sustainable development for different methods on this challenging benchmark. Moreover, current metrics such as AU-ROC have nearly reached saturation on simple datasets, which prevents a comprehensive evaluation of different methods. Inspired by the metrics in the segmentation field, we further propose several more practical threshold-dependent AD-specific metrics, ie, m$F_1$$^{.2}_{.8}$, mAcc$^{.2}_{.8}$, mIoU$^{.2}_{.8}$, and mIoU-max. Motivated by GAN inversion's high-quality reconstruction capability, we propose a simple but more powerful InvAD framework to achieve high-quality feature reconstruction. Our method improves the effectiveness of reconstruction-based methods on popular MVTec AD, VisA, and our newly proposed COCO-AD datasets under a multi-class unsupervised setting, where only a single detection model is trained to detect anomalies from different classes. Extensive ablation experiments have demonstrated the effectiveness of each component of our InvAD. Full codes and models are available at https://github.com/zhangzjn/ader.

Overview of Anomaly Detection Research under COCO-AD Benchmark

The paper introduces the concept of extending the well-known COCO dataset to create COCO-AD, a large-scale and general-purpose benchmark for multi-class anomaly detection (AD). This new benchmark aims to enhance the fair evaluation and development of anomaly detection methods across diverse scenarios, addressing the limitations of domain-specific datasets prevalent in industrial quality inspection and medical lesion examination.

Key Contributions

The research highlights three pivotal contributions to the field of visual unsupervised anomaly detection:

1. COCO-AD Dataset: The COCO-AD dataset is an innovative extension of the COCO dataset, designed specifically for anomaly detection. This dataset significantly increases the data scale and category diversity compared to existing AD datasets, allowing researchers to benchmark their algorithms in a complex and varied environment. It covers various semantic categories and mimics the challenging conditions encountered in real-world applications.
2. Novel Evaluation Metrics: To rectify the near-saturation of traditional metrics like AU-ROC on simple datasets, new AD-specific metrics such as mF$_1$^{.2}_{.8}$, mAcc$^{.2}_{.8}$, mIoU$^{.2}_{.8}$, and mIoU-max are proposed. These metrics offer practical, threshold-dependent evaluations, aligning more closely with real-world requirements in anomaly detection tasks.
3. InvAD Framework: The paper proposes the InvAD framework, an enhancement of reconstruction-based methods. Drawing from GAN inversion's powerful reconstruction capabilities, InvAD introduces high-precision feature reconstruction using spatial style modulation. This improves anomaly localization without complex augmentations or additional datasets.

Methodological Insights

The paper categorizes existing AD methods into three types: augmentation-based, embedding-based, and reconstruction-based. The InvAD framework falls under the reconstruction category, emphasizing simplicity and effectiveness through high-quality feature reconstruction. It leverages a novel approach inspired by learning-based GAN inversion, allowing for efficient and accurate anomaly detection under multi-class scenarios.

Experiments and Results

The efficacy of the proposed methods is evaluated using the COCO-AD dataset alongside traditional datasets like MVTec AD and VisA. InvAD consistently achieves better results, demonstrating its ability to overcome traditional limitations and adapt to the complex, large-scale nature of COCO-AD.

The results on COCO-AD particularly highlight its potential as a challenging and comprehensive benchmark, pushing the boundaries of what current methods can achieve and encouraging the development of more robust solutions.

Implications and Future Directions

The introduction of a general-purpose dataset like COCO-AD sets a new standard in anomaly detection research, prompting a shift towards more universal and challenging evaluations. The new metrics motivate solutions that align with practical industry needs, potentially leading to more deployable anomaly detection systems.

Further developments could explore optimizing InvAD with newer backbone architectures or integrating it with state-of-the-art techniques in other areas like natural language processing. As the field advances, focusing on enhancing computational efficiency and reducing resource consumption will be crucial for practical applications.

In summary, the research paper's introduction of COCO-AD and InvAD signifies a meaningful progression in anomaly detection, balancing theoretical advancements with practical implications. This work not only opens new avenues for researchers but also sets the stage for the application of anomaly detection in varied and complex real-world scenarios.