This was the final project for the subject digital image and video processing. Based on a 2000 image database (UKentuckyDatabase) I created an algorithm where you give it an image and it gives you an array of the most similar ones. In the database there are 4 "similar images" per image(including the original). So the F score is calculated that way

Author: Guillem Moreno
Date: May 2025

This project implements a content-based image retrieval (CBIR) system that compares image similarity using color-based descriptors. It builds upon a grayscale-based system (sistema1) and introduces a color descriptor using HMMD (Hue, Max-Min, and Diff).

The system (sistema2) consists of the following main stages:

• Input: RGB image of shape M × N × 3
• Output: HMMD color space using rgb2hmmd(I)

• Supported bins: 32, 64, 128, 256
• Uses predefined tables for hue and sum bins

• Normalized histogram vector (unit sum)

• Matrix H_db stores all descriptors: Ndb × numCells

• Distance metrics used:
  • Chi-squared
  • Bhattacharyya
  • Jensen-Shannon
  • Hellinger
  • KL-divergence

Each method was evaluated with different bin counts. Sample visualizations:

Bins	Precision-Recall Curve
256
128
64
32

• Hellinger with 64 bins: Best tradeoff between quality (F1=0.812) and speed (<1s per query)

• Main.m — system orchestration
• algo1_indexar.m — descriptor construction
• algo2_similitud.m — similarity calculation
• algo3_busqueda.m — query execution
• precision_recall.m — evaluation script
• rgb2hmmd.m — RGB to HMMD conversion
• quantizeHMMD.m — quantization logic
• you can also generate a random input.txt based on the generate_input_file.m

• MATLAB
• Image database: UKBench (ukbench%05d.jpg)
• Input: input.txt file with query image names