AnomalyXFusion: Multi-modal Anomaly Synthesis with Diffusion

TL;DR

The paper addresses the scarcity of abnormal samples in industrial anomaly localization and segmentation. It introduces AnomalyXFusion, a diffusion-based framework that fuses image, text, and mask modalities into an X-Embedding via a Multi-modal In-Fusion module and controls generation through a Dynamic Dif-Fusion module, with timestep-conditioned embeddings and a loss L that guides denoising. It also contributes MVTec Caption, adding 2.2k image-mask-text captions to enable semantic descriptions of anomalies, including logical ones. Empirically, AnomalyXFusion achieves higher fidelity and diversity in synthesized anomalies and improves anomaly detection/localization compared with state-of-the-art methods, demonstrating the value of multi-modal diffusion for industrial defect synthesis and perception.

Abstract

Anomaly synthesis is one of the effective methods to augment abnormal samples for training. However, current anomaly synthesis methods predominantly rely on texture information as input, which limits the fidelity of synthesized abnormal samples. Because texture information is insufficient to correctly depict the pattern of anomalies, especially for logical anomalies. To surmount this obstacle, we present the AnomalyXFusion framework, designed to harness multi-modality information to enhance the quality of synthesized abnormal samples. The AnomalyXFusion framework comprises two distinct yet synergistic modules: the Multi-modal In-Fusion (MIF) module and the Dynamic Dif-Fusion (DDF) module. The MIF module refines modality alignment by aggregating and integrating various modality features into a unified embedding space, termed X-embedding, which includes image, text, and mask features. Concurrently, the DDF module facilitates controlled generation through an adaptive adjustment of X-embedding conditioned on the diffusion steps. In addition, to reveal the multi-modality representational power of AnomalyXFusion, we propose a new dataset, called MVTec Caption. More precisely, MVTec Caption extends 2.2k accurate image-mask-text annotations for the MVTec AD and LOCO datasets. Comprehensive evaluations demonstrate the effectiveness of AnomalyXFusion, especially regarding the fidelity and diversity for logical anomalies. Project page: http:github.com/hujiecpp/MVTec-Caption

Figures (5)

• Figure 1: Framework of AnomalyXFusion. AnomalyXFusion aims to integrate multi-modal data into the process of anomaly synthesis. Specifically, the Multi-modal In-Fusion (MIF) module initially fuses various data modalities to form X-embeddings. Subsequently, the embeddings are dynamically adjusted based on the diffusion steps via the Dynamic Dif-Fusion (DDF) module.
• Figure 2: MVTec Caption Dataset. We present the MVTec Caption dataset, an enhancement to the existing MVTec AD and LOCO datasets, which now includes precise textual annotations for anomalies. These detailed descriptions not only resolve semantic ambiguities inherent in image-mask annotations but also facilitate the identification of logical anomalies. For instance, in the cable category, the annotations can clarify color errors, determine whether wires are missing or damaged; in the breakfast box category, they can specify incorrect counts of fruits or the presence of mixed grains. Without such precise semantic information, these distinctions would be challenging to ascertain.
• Figure 3: Analysis of Attributions in the Proposed MVTec Caption Dataset. As illustrated in the figure, the dataset encompasses a diverse array of textual keywords, which contributes to the diversity.
• Figure 4: Qualitative Evaluation. The qualitative assessment is facilitated by visualizing the output images, which are organized in sets from left to right as follows: the original anomaly-free image, the generated image with induced defects, the heatmap indicating the location of the defect generation, and the corresponding textual description.
• Figure 5: Visualization of component impacts. The results reveal that without the incorporation of text information, the attributes of the generated defects could not be precisely controlled. Upon introducing text cues, the generated defects began to reflect the text input, yet the quality of the defects remained suboptimal. However, with the integration of all components, we achieved the generation of high-quality defects with a high degree of controllability.