Evaluating Vision Transformer Models for Visual Quality Control in Industrial Manufacturing

Abstract: One of the most promising use-cases for machine learning in industrial manufacturing is the early detection of defective products using a quality control system. Such a system can save costs and reduces human errors due to the monotonous nature of visual inspections. Today, a rich body of research exists which employs machine learning methods to identify rare defective products in unbalanced visual quality control datasets. These methods typically rely on two components: A visual backbone to capture the features of the input image and an anomaly detection algorithm that decides if these features are within an expected distribution. With the rise of transformer architecture as visual backbones of choice, there exists now a great variety of different combinations of these two components, ranging all along the trade-off between detection quality and inference time. Facing this variety, practitioners in the field often have to spend a considerable amount of time on researching the right combination for their use-case at hand. Our contribution is to help practitioners with this choice by reviewing and evaluating current vision transformer models together with anomaly detection methods. For this, we chose SotA models of both disciplines, combined them and evaluated them towards the goal of having small, fast and efficient anomaly detection models suitable for industrial manufacturing. We evaluated the results of our experiments on the well-known MVTecAD and BTAD datasets. Moreover, we give guidelines for choosing a suitable model architecture for a quality control system in practice, considering given use-case and hardware constraints.

• The paper found that hierarchical Vision Transformer models like EsViT and EfficientFormer are suitable for real-world manufacturing due to their balance of performance, reduced size, and inference speed.
• The study shows that combining these transformers with representation-based anomaly detection methods, such as normalizing flows and Gaussian Mixture Models, yields promising results in defect detection and localization.
• The evaluated Vision Transformer methods achieved competitive performance compared to existing benchmark models on standard industrial anomaly detection datasets like MVTecAD.

Evaluating Vision Transformer Models for Visual Quality Control in Industrial Manufacturing

The integration of machine learning techniques into industrial manufacturing processes is a promising avenue for achieving early detection of defective products. This potential is realized through the deployment of advanced visual quality control systems, which help in minimizing production costs and reducing human errors associated with monotonous inspection tasks. The paper "Evaluating Vision Transformer Models for Visual Quality Control in Industrial Manufacturing" investigates the applicability and performance of Vision Transformer (ViT) models, integrated with anomaly detection algorithms, for enhancing visual quality control tasks in manufacturing settings.

Background and Motivation

Anomaly detection (AD) constitutes a critical aspect of visual quality control, aiming to identify rare and defective products in largely unbalanced datasets. Typically, an AD system comprises two pivotal components: a visual backbone for feature extraction and an anomaly detection algorithm that assesses the normalcy of these features. With the advent of transformer architectures, which offer superior capabilities in capturing global dependencies compared to conventional CNN-based backbones, it becomes vital to evaluate their potential in industrial applications. This research focuses on evaluating different combinations of state-of-the-art (SotA) vision transformer models and anomaly detection techniques to cater to industrial needs that demand efficiency in terms of model size, speed, and computational resource usage.

Methodology

The study examines state-of-the-art hierarchical vision transformer models and combines them with leading anomaly detection methods. By conducting a thorough assessment of these integrations, the research provides insights into the suitability of these models for real-world manufacturing scenarios. The evaluation encompasses the use of established datasets, such as MVTecAD and BTAD, which are widely recognized in industrial anomaly detection research. The authors also present guidelines for selecting the appropriate model architectures based on specific use-cases and hardware limitations.

Key Insights and Results

1. Performance of Vision Transformer Models: The paper reports that hierarchical transformers, such as EsViT and EfficientFormer, which offer reduced computational demands, are suitable candidates for real-world applications due to their ability to maintain performance while decreasing model size and inference time.
2. Anomaly Detection Techniques: Representation-based methods, particularly those leveraging normalizing flows (NFs) and Gaussian Mixture Models (GMMs), offer robust means for AD by estimating distributions of normal features and effectively identifying outliers. The results indicate promising AUROC scores in both detection and localization tasks, showcasing the potential of these approaches for accurate anomaly identification.
3. Comparison with Benchmark Models: When benchmarked against existing models, the proposed methods demonstrate competitive performance, achieving high scores in anomaly detection on datasets like MVTecAD. This reinforces the viability of employing transformers in quality control applications.

Practical Implications

The study’s outcomes have crucial implications for industrial settings. Employing vision transformers in conjunction with efficient AD techniques can result in highly scalable and accurate quality control systems. These systems facilitate the economic and reliable detection of defects, thus fostering improved product quality while also driving down labor costs.

Future Directions

The research opens avenues for future exploration in refining transformer-based models for better anomaly localization and real-time application. Further studies might focus on enhancing the scalability of these models to accommodate higher resolution images and various industrial contexts. Another potential path could involve hybrid architectures that combine the strengths of CNNs and transformers for optimized performance across diverse manufacturing tasks.

In conclusion, this paper contributes significant insights into the utility of vision transformers for industrial anomaly detection, offering a pathway to more automated and efficient quality inspection processes in manufacturing environments.