Improving Vision Anomaly Detection with the Guidance of Language Modality

TL;DR

This paper introduces a multimodal approach combining vision and language modalities to improve anomaly detection by reducing redundant information and learning a more meaningful latent space, leading to significant performance gains.

Contribution

The paper proposes Cross-modal Guidance (CMG) with CMER and CMLE to enhance vision anomaly detection by leveraging language modality for better feature focus and latent space learning.

Findings

CMG outperforms baseline by 16.81% in anomaly detection.

CMER effectively masks irrelevant image parts based on text.

CMLE creates a semantically meaningful latent space for vision data.

Abstract

Recent years have seen a surge of interest in anomaly detection for tackling industrial defect detection, event detection, etc. However, existing unsupervised anomaly detectors, particularly those for the vision modality, face significant challenges due to redundant information and sparse latent space. Conversely, the language modality performs well due to its relatively single data. This paper tackles the aforementioned challenges for vision modality from a multimodal point of view. Specifically, we propose Cross-modal Guidance (CMG), which consists of Cross-modal Entropy Reduction (CMER) and Cross-modal Linear Embedding (CMLE), to tackle the redundant information issue and sparse space issue, respectively. CMER masks parts of the raw image and computes the matching score with the text. Then, CMER discards irrelevant pixels to make the detector focus on critical contents. To learn a more compact latent space for the vision anomaly detector, CMLE learns a correlation structure matrix from the language modality, and then the latent space of vision modality will be learned with the guidance of the matrix. Thereafter, the vision latent space will get semantically similar images closer. Extensive experiments demonstrate the effectiveness of the proposed methods. Particularly, CMG outperforms the baseline that only uses images by 16.81%. Ablation experiments further confirm the synergy among the proposed methods, as each component depends on the other to achieve optimal performance.