Reconsidering learnable fine-grained text prompts for few-shot anomaly detection in visual-language models

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Control

Background:

• Few-Shot Anomaly Detection (FSAD) is critical for industrial image analysis with limited training samples.
• Current FSAD methods often rely on extensive manual prompt engineering for visual-language models, limiting accuracy and generalization.
• Domain gaps between training and testing datasets hinder the performance of text prompts in cross-dataset detection.

Purpose of the Study:

• To develop a unified framework for industrial FSAD using fine-grained learnable text prompts.
• To improve the accuracy and robustness of anomaly detection in industrial images.
• To overcome limitations of manual prompt engineering and domain shift in FSAD.

Main Methods:

• Proposed a Fine-grained Text Prompts Adapter (FTPA) with a registration loss to optimize text prompts for capturing fine-grained semantic information.
• Introduced a Dynamic Modulation Mechanism (DMM) to adaptively modulate image and text prompt branches, mitigating post-training errors and domain agnostic issues.
• Integrated FTPA and DMM into a visual-language model framework for enhanced FSAD.

Main Results:

• Achieved state-of-the-art performance in few-shot industrial anomaly detection and segmentation.
• Demonstrated high accuracy with AUROC scores of 98.3% for classification and 96.3% for segmentation in 4-shot settings on MVTec-AD.
• Achieved 93.8% AUROC for classification and 97.9% for segmentation in 4-shot settings on VisA dataset.

Conclusions:

• The proposed fine-grained learnable text prompt framework significantly enhances FSAD performance in industrial applications.
• The FTPA and DMM effectively address challenges of limited data, manual prompt limitations, and domain generalization.
• The method offers a robust and accurate solution for industrial anomaly detection and segmentation tasks.