README.md

metadata

title: Anomaly Detection
emoji: 📈
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 5.46.1
app_file: app.py
pinned: false

To use EfficientNet in the same way as ResNet-50 + PatchCore, just swap the backbone and pick an appropriate intermediate feature map. The steps are identical:

1. Backbone Selection & Freezing
   – Choose an EfficientNet variant (e.g. B4 or B5) pretrained on ImageNet.
   – Remove (or ignore) the final classification head and freeze all weights.

2. Feature-Map Extraction
   – Identify a mid-level block whose spatial resolution is neither too coarse nor too fine (e.g. the output of MBConv block 4 or 5).
   – Pass each input image through EfficientNet and take that block’s tensor of shape $$[C,H,W]$$.

3. Flatten into Patch Embeddings
   – For each spatial location $$(i,j)$$, flatten the $$C$$-dim vector into a patch embedding.
   – Collect all embeddings from your normal training images into a large memory bank.

4. Memory Bank & k-NN Detector
   – Fit a k-NN model (PatchCore) on the normal patch embeddings.
   – At test time, extract patch embeddings from new images and compute the Euclidean (or cosine) distance to nearest neighbors in the memory bank.
   – That distance is your anomaly score per patch.

5. Anomaly Map & Segmentation
   – Reshape the patch scores into an $$[H,W]$$ map.
   – Upsample to the original image resolution (bilinear) and apply a threshold or morphological filtering to segment abnormal regions.

Because EfficientNet’s inverted-bottleneck blocks are more parameter-efficient, you often get similar or better detection accuracy with lower FLOPs than ResNet-50—while the overall PatchCore workflow remains unchanged.