Multi-Task Representation Learning for Robust Multi-Class Anomaly Detection = 다중 클래스에서 강인한 이상 탐지를 위한 다중 태스크 표현 학습

다국어 초록 (Multilingual Abstract)

In this work, we demonstrate that using Multi-Task Representation Learning to capture both pixel-level details and high-level semantics improves anomaly detection performance.
This study proposes a multi-task learning (MTL)-based approach to address the challenge of applying supervised learning methods in image-based anomaly detection, where abnormal data are scarce. Existing multi-class unsupervised anomaly detection studies, such as ViTAD, share the common characteristic of utilizing a Vision Transformer encoder pretrained on large-scale datasets. However, these methods still exhibit limitations. In particular, ViTAD struggles with the reconstruction of fine-grained local defects and relatively low pixel-level segmentation performance. To overcome these issues, this study designs an MTL framework in which a shared encoder learns the detailed representations of normal data through a combination of reconstruction and classification tasks. The reconstruction task enables the encoder to learn fine-grained structural features at the pixel-level, while the classification task strengthens global semantic discrimination, all owing the encoder to simultaneously learn detailed texture information and high-level representations that more clearly define inter-class boundaries in the latent space. Through this MTL-based representation learning, the proposed model achieves more precise segmentation performance and improved pixel-level AP and F1 scores compared to ViTAD.

목차 (Table of Contents)

• Ⅰ. Introduction 1
• Ⅱ. Related Works 6
• 1. Multi-Class Unsupervised Anomaly Detection 6
• 2. Multi-Task Learning based Anomaly Detection 11
• Ⅲ. Proposed Method 14
• 1. Task Definition of Multi-Task Learning 14
• 2. Encoder Training Term 16
• 3. Harmonic Mean Loss 24
• 4. Decoder Training Term 27
• Ⅳ. Experimental Results & Analysis 31
• 1. Dataset 31
• 2. Evaluation Metrics 37
• 3. Experimental Settings 43
• 4. Quantitative Analysis 45
• 5. Qualitative Analysis 64
• 6. The Role of the Exponential in the Harmonic Mean Loss 81
• V. Conclusion & Future Works 84