This project aims to classify images of flowers from the Oxford 102 Flowers Dataset using clustering and classification techniques. The project is divided into two main phases: feature extraction using clustering and classification based on extracted features.
The dataset used in this project is the Oxford 102 Flower Dataset. It contains images of flowers from 102 different classes.
-
Extract Valuable Image Regions:
- Extract color and spatial features from each pixel.
- Perform clustering to identify regions containing flowers.
- Use K-means clustering algorithm with various parameters to optimize region extraction.
-
Evaluate Clustering:
- Ensure extracted regions have spatially close pixels and similar colors.
- Avoid extracting regions that are too small or too large.
- Adjust the importance of spatial and color features to optimize clustering.
- K-means Clustering: Applied to extract significant image regions.
- Parameter Tuning: Various algorithms and parameters were tested to find the best clustering configuration.
- Evaluation Metrics: Both qualitative and quantitative metrics (at least two) were used to evaluate the clustering performance.
-
Feature Extraction:
- Extract color statistical features and shape features from each region.
- Construct feature vectors for images by clustering all features and creating histograms of features per image.
-
Classification:
- Use the extracted feature vectors to train a classifier.
- Evaluate the model using metrics such as accuracy, precision, recall, and F1-score.
- Statistical and Shape Features: Extracted from identified regions.
- Feature Vector Construction: Histograms of clustered features were created for each image.
- Classifier Training: Trained a classification algorithm to predict flower classes based on feature vectors.
calculate_mean_imgs_clstrs_features: Computes the mean feature vector for clusters.calculate_mean_imgs_clstrs_except_index: Computes the mean feature vector excluding a specific cluster.cal_classify_results: Classifies images and outputs classification results.check_clusters_importance: Determines the importance of clusters by evaluating classification confidence changes when clusters are removed.remove_least_important_clusters: Removes the least important clusters from the feature set.
-
Feature Extraction:
- Extract features for each region and create histograms.
- Use clustering to identify and rank the importance of each region.
-
Classification:
- Train a classifier with the constructed feature vectors.
- Propose improvements based on error analysis.
The project successfully identified and classified flower images with notable efficiency. The key achievements include:
- Optimal Clustering Configuration: Through extensive testing and parameter tuning of the K-means clustering algorithm, we identified the best configuration that effectively extracted meaningful regions from the flower images.
- Feature Extraction: Extracted both color statistical and shape features from identified regions, which were then used to construct comprehensive feature vectors for each image.
- Cluster Analysis: Using custom functions such as
check_clusters_importanceandremove_least_important_clusters, we evaluated the importance of each cluster by measuring changes in classification confidence when clusters were removed. - Confidence Impact: We observed that removing the least important clusters led to a minimal decrease in classification confidence, which helped refine the feature set to focus on the most critical regions.
- Reduction in Feature Set: By removing the least important clusters, we reduced the feature set size by 20% without significantly impacting classification performance, resulting in a more efficient model.
- Visualization: Important clusters for each image were visualized, providing insights into which regions the classifier focused on for making decisions.
