https://github.com/kisaa-fatima/foreground-segmentation-using-kmeans-face-recognization-using-knn
Implement a basic version of the interactive image cut-out/segmentation approach called Lazy Snapping. The program uses K-Means Clustering to segment images into foreground and background based on user-provided seed pixels.
https://github.com/kisaa-fatima/foreground-segmentation-using-kmeans-face-recognization-using-knn
face-recognition foreground-segmentation kmeans-clustering knn-classification matplotlib numpy opencv2 pil
Last synced: 16 days ago
JSON representation
Implement a basic version of the interactive image cut-out/segmentation approach called Lazy Snapping. The program uses K-Means Clustering to segment images into foreground and background based on user-provided seed pixels.
- Host: GitHub
- URL: https://github.com/kisaa-fatima/foreground-segmentation-using-kmeans-face-recognization-using-knn
- Owner: Kisaa-Fatima
- Created: 2024-07-14T08:41:26.000Z (11 months ago)
- Default Branch: main
- Last Pushed: 2024-07-14T08:50:18.000Z (11 months ago)
- Last Synced: 2025-02-17T16:41:33.957Z (3 months ago)
- Topics: face-recognition, foreground-segmentation, kmeans-clustering, knn-classification, matplotlib, numpy, opencv2, pil
- Language: Jupyter Notebook
- Homepage:
- Size: 16.2 MB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Foreground-segmentation-using-kmeans-AND-Face-recognization-using-KNN
## Interactive Foreground Segmentation and Face Recognition
This repository contains two main ASSIGNMENTS:
- Interactive Foreground Segmentation Using K-Means Clustering
- Face Recognition Using K-Nearest Neighbors (K-NN)
Each ASSIGNMENT involves detailed data preprocessing, algorithm implementation, and result analysis. Below are the instructions and descriptions for each project.
## Assignment 1: Interactive Foreground Segmentation Using K-Means Clustering
## Objective
Implement a basic version of the interactive image cut-out/segmentation approach called Lazy Snapping. The program uses K-Means Clustering to segment images into foreground and background based on user-provided seed pixels.
## Libraries Used
time: To measure the processing time for segmentation.
PIL (Image): For opening and manipulating images.
matplotlib.pyplot (plt): To display the segmentation results.
numpy: For numerical operations.
cv2 (OpenCV): For image processing tasks like reading images, converting color spaces, and displaying images.
## Solution
- Segmentation Mask Processing:
Create three versions of the input image:
One with background pixels set to black.
One with foreground pixels set to black.
The original image.
- Lazy Snapping:
Utilize human annotations (foreground and background seed pixels) to compute a precise figure-ground segmentation.
- Optimization:
K-Means Optimization: Use Mini-batch K-Means for efficiency.
Reduce Color Space Dimensionality.
Parallelize K-Means computation for foreground and background classes.
Results and Observations
- N Values:
Increasing N results in more refined segmentation but increases computational complexity.
Lower N values provide coarse segmentation but reduce computation time.
- Comparison:
Varying N values and strokes impact segmentation consistency and robustness.
Running the Scripts
## Assignment 2: Face Recognition Using K-Nearest Neighbors (K-NN)
## Objective
Implement a K-Nearest Neighbors classifier from scratch for face recognition using the CMU Pose, Illumination, and Expression (PIE) Dataset.
## Libraries Used
- numpy: For numerical operations.
- matplotlib: For visualization.
## Tasks
- Pre-process the Dataset:
Normalize each face image vector to unit length.
Split the dataset into training and testing sets.
- Implement K-NN Classifier:
Use Euclidean and cosine similarity distance measures.
Evaluate performance for different k-values (2, 5, 7, 11).
- Apply SVM and GaussianNB:
Compare accuracy with K-NN.
Perform Dimensionality Reduction:
Visualize datasets in 3-D using PCA.
Preprocessing Steps
- Normalization:
Divide each vector by its magnitude.
- Splitting Dataset:
Randomly select 150 images for training and 20 for testing for each subject.
Results and Observations
- Accuracy Comparison:
- SVM: Highest accuracy (~97.65%).
- K-NN: Competitive accuracy (~96.08%).
GaussianNB: Lower accuracy (~75.69%).
- Distance Metrics:
- Euclidean Distance: Effective for face recognition.
- Cosine Distance: Less effective for this dataset.