Introduction

comparing features of target images with normal images

Untitled

— from Deep Industrial Image Anomaly Detection: A Survey. MIR24

  1. memory-based methods regularly do not require the loss function for training, and models are constructed quickly.
    1. just use pre-trained network.

Methods

  1. Spade, 20 image level memory bank of nominal patch features without downsampling, kNN
  2. PatchCore**, cvpr22** Spade with memory bank downsampling, kNN
    1. with better patch size, patch feature
    2. FAPM, ICASP23
      1. speed-up PatchCore by patch-wise memory bank & layer-wise memory bank;
      2. patch-wise memory bank: 假设image alignment，即把bank的建立和test时候的比较，都限定在patch level而不是image level的全局操作。
      3. adaptive coreset sampling
      4. 竟然63 citation了
    3. SA-PatchCore, ieee access 23
  3. CFA 是“参数化可训练”的特征映射，而 APG-Net 是“非参数化纯数学”的特征空间重塑。目的一样，方法不同。
    1. CFA, access22
      1. 竟然368 citation了
    2. APG-Net, tcsvt26
    3. 异同：
      1. APG-Net：EMA 用于“特征归一化（Feature Normalization）”。
        1. 它利用 EMA 来动态更新和记录训练集特征的全局均值（Mean）和方差（Variance）。通过这个EMA统计量，将特征拉到一个标准分布下，然后再去做原型匹配。EMA 的作用是稳定特征尺度。
      2. CFA：EMA 用于“记忆库的更新压缩（Memory Bank Update）”。
        1. CFA 为了控制 Memory Bank 的大小，不直接存储所有特征，而是用 EMA 来迭代更新记忆库中的聚类中心（Memorized Features）。新输入的特征会通过 EMA 与旧的记忆中心融合，从而实现与目标数据集大小无关的记忆库压缩。
    4. 总结
      1. APG-Net走的是纯数学空间重塑路线，通过 EMA 稳定全局统计量，用非参数化的向量引导公式直接在特征空间中“捏”出更具区分度的分布，主打零额外参数和极简部署。
      2. CFA走的是深度学习/迁移学习路线，通过 EMA 更新记忆库，用度量学习（Contrastive Loss）训练额外的网络层来适应目标数据。
  4. PNI, 23
    1. memory bank with trainable parameters
  5. MulSen-AD, cvpr25
    1. propose MulSen-TripleAD, a decision-level fusion algorithm that integrates these three modalities

Discussion

  1. image level memory bank of patch features, less dependent on matched locations, but no position & neighborhood info considered.
  3. Their performance is ensured by pre-trained networks and additional memory space,
  4. and this type of method is currently the most effective in IAD tasks????
  5. slow & high memory

PatchCore**, cvpr22**