CBIR

An academic Content-Based Image Retrieval implementation with a simplified HOG descriptor and spatial information. The included dataset has 25 documents across five classes and five queries.

Each document is split into nine regions. For every query, the system automatically selects the largest connected component region and compares its descriptor against the indexed regions.

1. Images in grayscale, 256 x 256.
2. Fixed 3 x 3 grid of regions.
3. Simplified 128-dimension HOG descriptor.
4. Cosine similarity combined with spatial IoU.
5. Top-5 ranking per query.

score = 0.8 * cosine_similarity + 0.2 * IoU

These numbers describe only the five included queries and should not be read as a general benchmark. The small dataset demonstrates the pipeline but is not enough to compare approaches with statistical significance.