Color Feature Extraction

Visual content relies heavily on color distribution, making it a primary metric for identifying similar images. Color histograms represent this distribution by mapping pixel frequencies across specific color channels, such as RGB or HSV. This multidimensional array effectively captures the global color composition without relying on spatial relationships.

Retrieval Workflow

The search mechanism involves four primary stages:

1. Vector Extraction: Compute the color histogram for both the source query and the target dataset.
2. Distance Measurement: Quantify the variance between the query vector and dataset vectors using metrics like Euclidean, Manhattan, or Chi-squared distances.
3. Ranking: Sort the dataset entries in ascending order based on the calculated distance; smaller distances indicate higher similarity.
4. Output: Retrieve the top-ranking matches.

Strengths and Limitations

Utilizing color vectors offers significant efficiency and computational speed. The approach is inherently invariant to geometric transformations like rotation and scaling. However, relying solely on global color distribution ignores spatial structure and semantic context, creating a gap between low-level pixel data and high-level human perception. Furthermore, high-dimensional histograms can degrade performance, and global color is highly susceptible to lighting variations and noise.

Optimization Strategies

To mitigate these limitations, several enhancements can be applied:

• Color Space Transformation: Converting RGB to HSV or CIELAB spaces separates chromaticity from intensity, improving robustness against illumination shifts.
• Spatial Segmentation: Dividing the image into localized blocks before extracting histograms preserves spatial relationships.
• Feature Weighting: Assigning relevance scores to specific color bins emphasizes dominant visual traits.
• Dimensionality Reduction: Applying machine learning models or PCA to compress the feature space while retaining discriminative power.

Practical Domains

This methodology supports diverse fields, including digital asset management, texture-based web search, automated categorization of scenes, and preliminary analysis in medical diagnostics.

Code Implementation

% Color-Based Image Search Engine
clc; clear;

totalImgs = 36;
dbPath = 'flora_db\';
queryFile = 'flora_db\query.jpg';

queryHist = extractColorHist(queryFile);

distMetrics = zeros(1, totalImgs);

for idx = 1:totalImgs
    dbFile = sprintf('%s%d.jpg', dbPath, idx);
    dbHist = extractColorHist(dbFile);
    
    % Compute Euclidean distance across 3 channels (4 bins each)
    diffSq = (queryHist - dbHist).^2;
    binDist = sqrt(sum(reshape(diffSq, 4, 3), 1));
    distMetrics(idx) = mean(binDist);
end

% Sort distances in ascending order
[sortedDists, sortIdx] = sort(distMetrics);

% Display query and top 3 matches
figure;
subplot(2, 2, 1);
imshow(imread(queryFile));
title('Source Query');

for rank = 1:3
    bestMatchPath = sprintf('%s%d.jpg', dbPath, sortIdx(rank));
    subplot(2, 2, rank + 1);
    imshow(imread(bestMatchPath));
    caption = sprintf('Match Rank %d - Dist: %.4f', rank, sortedDists(rank));
    title(caption);
end

function h = extractColorHist(imgPath)
    % Placeholder for histogram extraction logic
    % Returns a 12-element vector (3 channels x 4 bins)
    img = imread(imgPath);
    % ... implementation details ...
end