AdaptCLIP: Adapting CLIP for Universal Visual Anomaly Detection
Bin-Bin Gao, Yue Zhu, Jiangtao Yan, Yuezhi Cai, Weixi Zhang, Meng Wang, Jun Liu, Yong Liu, Lei Wang, Chengjie Wang
2025-05-16
Summary
This paper talks about AdaptCLIP, a new method that helps AI spot unusual or abnormal things in pictures, even if it hasn't seen examples of those specific problems before.
What's the problem?
The problem is that most AI systems need lots of training on specific types of weird or abnormal images to recognize them, which isn't always possible, especially when new or rare problems pop up in different situations.
What's the solution?
The researchers created AdaptCLIP, which uses a smart way of comparing and adapting what the AI knows about images, so it can quickly notice when something looks out of place, even with little or no extra training on those new types of images.
Why it matters?
This matters because it means AI can be used to catch problems early in areas like medicine, security, or manufacturing, even when those problems are new or rare, making these systems more flexible and reliable.
Abstract
AdaptCLIP uses adaptive representations and comparative learning to achieve zero-/few-shot anomaly detection across domains without additional fine-tuning.