The recent developments in the field of vision-language models (VLMs) and anomaly detection (AD) showcase a significant push towards enhancing generalization, robustness, and unsupervised learning capabilities. Innovations are particularly focused on overcoming limitations related to overfitting, noisy data, and the challenge of selecting optimal models for downstream tasks without labeled data. Techniques such as Mixture-of-Prompts Distillation (MoPD) and Visual-tExtual Graph Alignment (VEGA) are at the forefront of improving VLMs' adaptability and performance on unseen classes and tasks. In the realm of AD, advancements like SoftPatch+ and Cross-modal Normality Constraint (CNC) are addressing the critical issue of noisy training data and over-generalization in decoders, respectively, thereby setting new benchmarks for unsupervised anomaly classification and segmentation.
Noteworthy papers include:
- MoPD: Mixture-of-Prompts Distillation for Vision-Language Models: Introduces a novel method to enhance the generalization ability of soft prompts on unseen classes by transferring knowledge from hard prompts.
- Enhancing Table Recognition with Vision LLMs: Proposes a Neighbor-Guided Toolchain Reasoner (NGTR) framework that significantly improves table recognition capabilities by addressing low-quality image input issues.
- Learning to Rank Pre-trained Vision-Language Models for Downstream Tasks: Presents VEGA, a method for unsupervised selection of VLMs based on the alignment of visual and textual features, offering a reliable solution for unlabeled downstream tasks.
- SoftPatch+: A fully unsupervised anomaly detection method that effectively denoises data at the patch level, demonstrating robust performance in real-world industrial inspection scenarios.
- CNC: Cross-modal Normality Constraint for Unsupervised Multi-class Anomaly Detection: Introduces a novel approach to mitigate decoder over-generalization in multi-class anomaly detection, leveraging class-agnostic learnable prompts for improved performance.
