VisA 工业视觉异常检测数据集

VisA 数据集是用于异常检测和分割的 SPot-the-Difference 自监督预训练数据集。包含 12 个子集，对应 12 个不同的对象，如图所示。有 10,821 张图像，其中 9,621 个正常样本和 1,200 个异常样本。四个子集是不同类型的印刷电路板 (PCB），其结构相对复杂，包含晶体管、电容器、芯片等。对于视图中多个实例的情况，我们收集四个子集：Capsules 、 Candles 、 Macaroni1 和 Macaroni2 。 Capsules 和 Macaroni2 中的实例在位置和姿势上有很大不同。此外，研究团队还收集了四个子集，包括腰果、口香糖、炸薯条和管道炸薯条，其中对象大致对齐。异常图像包含各种缺陷，包括划痕、凹痕、色斑或裂纹等表面缺陷，以及错位或缺失部件等结构缺陷。

The VisA dataset is a SPot-the-Difference self-supervised pre-training dataset designed for anomaly detection and segmentation. It consists of 12 subsets corresponding to 12 distinct object types, as illustrated in the figure. In total, the dataset contains 10,821 images, including 9,621 normal samples and 1,200 anomalous samples. Four of the subsets cover different types of Printed Circuit Boards (PCBs), which have relatively complex structures with components including transistors, capacitors, and chips. For cases involving multiple instances in a single view, four subsets are collected: Capsules, Candles, Macaroni1, and Macaroni2. The instances in Capsules and Macaroni2 vary considerably in terms of their positions and poses. Additionally, the research team has collected another four subsets including Cashews, Gum, French Fries, and Tube Fries, where the objects are roughly aligned. Anomalous images feature a diverse array of defects, including surface defects such as scratches, dents, discolorations, or cracks, as well as structural defects like misplaced components or missing parts.

搜集汇总

背景与挑战

背景概述

VisA数据集是一个用于异常检测和分割的自监督预训练数据集，专注于工业视觉应用。它包含12个子集，覆盖印刷电路板、胶囊等多种对象，总计10,821张图像，其中异常样本1,200张，涵盖划痕、错位等多种缺陷类型。该数据集旨在支持工业场景下的视觉异常识别和分割任务。