This repository contains a CBIR (content-based image retrieval) system Implimentation for my conference paper
Extract query image's feature, and retrieve similar ones from image database
In this system our approach is based on the ViT architecture for ex- tracting features from the Corel images. Then, we used the PCA as a features selector to minimize the dimensionality. Finally, we implemented the Annoy algorithm for similarity searches.
- Vision Architecture
The curse of dimensionality told that vectors in high dimension will sometimes lose distance property
- PCA
CBIR system retrieves images based on feature similarity
we have evaluated our CBIR system using the Cored-1K dataset with the same ten aforementioned classes; then, the precision and recall metrics are computed for each category, individually and overall. These metrics are calculated based on the top 20 images retrieved. Table below shows the findings results.
| Category | Precision (%) | Recall (%) |
|---|---|---|
| Buses | 100 | 20 |
| Mountains | 100 | 20 |
| Beach | 90 | 18 |
| Elephants | 100 | 20 |
| Food | 100 | 20 |
| Flowers | 100 | 20 |
| Africa | 100 | 20 |
| Horses | 100 | 20 |
| Dinosaurs | 100 | 20 |
| Buildings | 100 | 20 |
| Average | 99 | 19.8 |
Zitane Smail / @zitansmail