Content-Based Image Retrieval using Color Histograms

Keyword search fails the moment images have no reliable tags — personal photo libraries, surveillance archives and product catalogues all hold millions of untagged pictures. Content-based image retrieval (CBIR) solves this by describing each image with numbers computed from its own pixels and ranking the collection by visual similarity to a query image. This project implements the classical, fully transparent CBIR pipeline end to end: every image is reduced to a 2×2 spatial RGB histogram (256 dimensions) plus Sobel edge density (1 dimension), the vectors are L2-normalized and indexed, and a query is answered by chi-squared distance ranking. A live web demo makes the engine tangible — pick any photo as the query and the app shows the ranked results with similarity scores, while a feature inspector reveals the exact 257-dimensional vector the engine computed. Because every stage is classical and inspectable, the student can explain precisely why each result was returned.

1. Each collection image is divided into a 2×2 spatial grid; a 64-bin RGB histogram is computed per cell and concatenated into a 256-D vector.
2. Sobel edge density (fraction of edge pixels) is appended, forming the final 257-D descriptor, which is L2-normalized and stored in the index.
3. A query image goes through the identical extraction, guaranteeing query and index vectors live in the same space.
4. Chi-squared distance is computed between the query vector and every indexed vector; the collection is ranked by ascending distance and the top-K returned.
5. The evaluation harness replays every benchmark image as a query, marks same-category hits as relevant, and computes Precision@K and mAP.
6. The fused variant concatenates the 257-D classical vector with a 2048-D CNN embedding before ranking, and the harness measures the improvement.

Tech stack:

• Python 3, NumPy, OpenCV (feature extraction)
• scikit-learn (metrics)
• Jupyter Notebook (evaluation harness)
• Flask/FastAPI (REST query API)
• HTML5 + JavaScript (live retrieval demo)
• Wang image database (Corel-1K)

Dataset & model details

• Dataset: Wang image database (Corel-1K, James Z. Wang, Penn State) — 1,000 photographs in 10 categories × 100 images (African people, beaches, buildings, buses, dinosaurs, elephants, flowers, horses, mountains, food). Ground truth: same-category images count as relevant for a query.
• Task: Query-by-example retrieval; input = a query image, output = the collection ranked by visual similarity (top-K list with similarity scores).
• Model: Classical: 257-D descriptor (2×2 spatial RGB histogram + Sobel edge density), L2-normalized, chi-squared distance ranking. Fused variant: 257-D classical concatenated with a 2048-D CNN embedding, cosine/chi-squared ranking.
• Metrics: Precision@10 ≥ 0.72 (design target), mAP ≥ 0.65 (design target), Precision@K curves for K = 5/10/20/50, query latency < 300 ms per 1,000 images (expected). No metric is claimed as measured until the evaluation harness runs for the order.

• [Real histogram-matching engine] 2×2 spatial RGB histograms plus Sobel edge density extracted live from actual pixels — 257-D vectors, L2-normalized, ranked by chi-squared distance. No tags, no filenames, pure visual similarity.
• [Live retrieval demo] Pick any photo as the query; the engine extracts its features and returns the collection ranked most-similar-first with similarity scores and rank badges.
• [Feature inspector] Select any image to see its exact RGB histograms (16 bins/channel), edge density and vector norm — the full internals of the engine, exposed.
• [Pairwise distance matrix] The complete chi-squared distance matrix over the demo collection, showing how the metric separates similar from dissimilar images.
• [Evaluation harness] Precision@K and mean average precision computed on the Wang (Corel-1K) benchmark with same-category ground truth, in a reproducible notebook.
• [Deep-fusion variant] The classical 257-D descriptor concatenated with a 2048-D CNN embedding for a fused representation, with the measured gain documented.
• [Web UI + REST API] The retrieval engine served behind a clean web interface and a query API, as the pipeline diagram in the demo shows.

• Complete CBIR engine (ingest, feature extraction, indexing, ranking, serving)
• Evaluation harness notebook with Precision@K and mAP on Corel-1K
• Live retrieval + feature inspector + system report web demo
• Pairwise distance matrix and precision@K-by-feature-set charts
• Project report PDF (retrieval theory, histogram methods, distance metrics, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (color spaces, histograms, chi-squared, precision/recall in retrieval)

• Color histograms ignore spatial layout within each grid cell and object semantics — a red car and a red apple can rank as similar; the report discusses this honestly.
• Precision@10 ≥ 0.72 and mAP ≥ 0.65 are design targets for the evaluation run, stated honestly — the report documents the actual measured figures after the harness runs.
• The demo collection is a small curated set for interactivity; the full 1,000-image benchmark runs in the evaluation notebook, not in the browser.
• The fused CNN variant improves semantic matching but adds inference cost — the tradeoff is measured and reported.
• Retrieval quality depends on the collection: near-duplicate or very dark images compress the distance range.