Skip to main content

​
Introduction

As we discussed in the pretrained CNN section, the aim is to bring the high-dimensional vector data from the CNN in a lower dd-dimensional space (a hyperparameter) and to use the dd-dim features for the downstream tasks. To optimize the hyperparameter of what should be the best dd-dim space we use kNN to classify the images using the accuracy of the classification as a metric for the adopted grid search optimization approach. Typically kNN is implemented inside a Qdrant vector database but the analysis uses an exact kNN algorithm without any dependency on the vector database.

​
Visualizing Embeddings

Please note that in the plots a lexicographic mapping is used to map the labels (PASS and FAIL) to the colors.
The UMAP algorithm reduces the 2048-dimensional ResNet-50 embeddings to 2D and 3D for visualization purposes. These visualizations show the separation between nominal (PASS) and anomalous (FAIL) samples in the embedded space.

​
UMAP kNN Results

The kNN classifier is applied to the UMAP-reduced embeddings with various values of kk and embedding dimensions dd. The results show:
  • Best Performance: AU-ROC of approximately 0.97 achieved with UMAP dimensionality reduction paired with kNN using majority voting
  • Optimal kk: Values between 3-7 neighbors typically perform best
  • Optimal dd: Embedding dimensions between 10-50 provide good discrimination while maintaining computational efficiency
The high AU-ROC score indicates that the pretrained ResNet-50 features, when projected via UMAP and classified with kNN, can effectively distinguish between nominal and anomalous product seam images despite the visual similarity of the raw images.
Edit this page on GitHub or file an issue.
Connect these docs to Claude, VSCode, and more via MCP for real-time answers.