Advanced Computer Vision and Video Analytics

A comprehensive implementation of advanced computer vision techniques and video analytics solutions using Python, OpenCV, and deep learning frameworks.

Table of Contents

• Overview
• Features
• Installation
• Project Structure
• Modules & Implementations
• Notebook: Image Processing & Digits Classification
• Notebook: Shape and Image Transformations
• Notebook: Edge Detection and Image Segmentation
• Notebook: Histogram Analysis, Equalization & DFT
• Notebook: Image Compression & Deep Learning Classification
• Notebook: Segmentation, Detection & Classification
• Notebook: Blob Detection, Image Enhancement & Classification
• Notebook: SIFT, ORB, Watershed, ResNet & Few-Shot Learning
• Notebook: Stitching, Denoising, GANs & Segmentation Playground
• Notebook: Image Denoising & Video Action Pipeline

Overview

This repository contains implementations of advanced computer vision algorithms and video analytics techniques including:

• Image transformations and geometric operations
• Object detection and tracking
• Face recognition and analysis
• Pose estimation
• Video processing and frame analysis
• Motion detection and activity recognition
• Semantic and instance segmentation
• 3D reconstruction

Features

✨ Image Processing

• Geometric transformations (rotation, scaling, translation, shearing, reflection)
• Image filtering and enhancement
• Morphological operations
• Edge detection and contour analysis

✨ Video Analytics

• Real-time video processing
• Frame-level analysis
• Motion tracking
• Temporal analysis

✨ Deep Learning Integration

• Pre-trained model support (YOLO, ResNet, etc.)
• Custom model implementations
• Transfer learning examples

✨ Visualization Tools

• Matplotlib-based visualization
• Real-time plotting
• Annotated frame display

Installation

Prerequisites

• Python 3.8+
• pip or conda

Setup

git clone https://github.com/anshika1279/Computer-Vision-Implementation.git
cd Computer-Vision-Implementation
pip install -r requirements.txt

Key runtime dependencies include:

• Notebook tooling: nbformat, nbconvert, notebook (for cleaning/fixing metadata)
• CV/DL extras: timm (MobileNetV1), ultralytics, torch/torchvision, tensorflow

Project Structure

• README.md: Project overview and instructions.
• requirements.txt: Python dependencies for all notebooks.
• image_processing_and_digits_classification.ipynb: Image resizing/blur demo plus digits classification with multiple models.
• ShapeAndImageTransformations.ipynb: Shape and image transformation examples.
• edge_detection_and_image_segmentation.ipynb: Edge detection operators and image segmentation techniques.
• Histogram_Analysis_Equalization_DFT.ipynb: Histogram analysis, contrast enhancement, and frequency domain transformations.
• image_compression_techniques_DCT_Deep_learning_image_classification.ipynb: DCT compression and CNN-based digit/object classification.
• segmentation_detection_classification.ipynb: Advanced CV pipeline with edge/region segmentation, Hough transform, YOLO/R-CNN detection, and Fashion-MNIST/CIFAR-100 classification.
• blob_detection_image_enhancement_classification.ipynb: Blob detection algorithms (LoG, DoG, DoH), comprehensive image enhancement techniques, and transfer learning with AlexNet/VGG16 on CIFAR-100.
• sift_orb_watershed_resnet_few_shot_learning.ipynb: Feature detection (SIFT, ORB), feature matching (BFMatcher), watershed segmentation, ResNet-18/34 classification on CIFAR-100, and few-shot/one-shot learning with elastic deformation augmentation.
• Stitching_Denoising_GAN_SegmentationPlayground.ipynb: Colab-ready playground covering image stitching (simple/panorama/ORB + pose), inpainting, MNIST denoising autoencoders, GANs (MNIST, CIFAR-10), MobileNet V1/V2/V3 fine-tuning, and notebook metadata fixes for GitHub rendering.
• Image_Denoising_Video_Action_Pipeline.ipynb: Image denoising comparison (Median, Wavelet, Noise2Void U-Net) and video action recognition pipeline (frame extraction, video processing, UCF101 subset with 3D CNN classification).

Modules & Implementations

• Image resizing with multiple interpolation methods and blurring with box, Gaussian, and bilateral filters.
• Digits classification using sklearn digits dataset with Gaussian Naive Bayes, RBF SVM, and Random Forest, including cross-validation and ROC visualization.

Notebook: Image Processing & Digits Classification

• File: image_processing_and_digits_classification.ipynb
• Part 1: Resize (linear, nearest, cubic) and blur (box, Gaussian, bilateral) a local image; expects image.png in the repo root and displays a comparison grid.
• Part 2: Train/evaluate classifiers (Gaussian Naive Bayes, RBF SVM, Random Forest) on sklearn digits with 5-fold CV; prints metrics and shows ROC curves.

Notebook: Shape and Image Transformations

• File: ShapeAndImageTransformations.ipynb
• Part 1: 2D rectangle transformations (translate, scale, rotate, reflect, shear, composite) visualized with Matplotlib.
• Part 2: Image transformations on input.jpg using OpenCV (translate, reflect, rotate, scale, crop, shear on x/y) with side-by-side plots.
• Part 3: Additional 2D shape transformations with reusable helpers for translate/scale/rotate/reflect/shear and composite examples.

Notebook: Edge Detection and Image Segmentation

• File: edge_detection_and_image_segmentation.ipynb
• Edge Detection: Implements multiple edge detection operators including:
  • Sobel operator (combined X and Y gradients)
  • Prewitt edge detection
  • Roberts cross operator
  • Canny edge detector
• Image Segmentation: Demonstrates various segmentation techniques:
  • Global thresholding (fixed threshold binarization)
  • Adaptive thresholding (local neighborhood-based)
  • Watershed segmentation with morphological operations
• Preprocessing: Includes color space conversions (BGR→RGB→Grayscale→Binary) and image metrics calculation
• Visualization: Displays all results in a comprehensive grid layout with labeled subplots
• Outputs: Saves processed images (edge maps, segmented regions) for further analysis

Notebook: Histogram Analysis, Equalization & DFT

• File: Histogram_Analysis_Equalization_DFT.ipynb
• Histogram Analysis: Computes and plots histograms for both grayscale and color (RGB) images
  • Individual channel histograms for color images (B, G, R)
  • Histogram normalization to probability distributions
  • Visualization with matplotlib for histogram analysis
• Contrast Enhancement: Implements histogram equalization for improving image contrast
  • Before/after comparison of grayscale images
  • Visual quality assessment with text annotations
  • Side-by-side display of original and equalized images
• Discrete Fourier Transform (DFT): Frequency domain analysis and transformations
  • DFT computation with magnitude spectrum visualization
  • Inverse DFT for image reconstruction
  • Rotation property verification (45° rotation test)
  • Demonstrates spatial vs. frequency domain correspondence
• Compatible with Google Colab: Uses cv2_imshow for Colab environments

Notebook: Image Compression & Deep Learning Classification

• File: image_compression_techniques_DCT_Deep_learning_image_classification .ipynb
• DCT-Based Image Compression: Implements both lossy and lossless compression techniques
  • Lossy compression with quantization using JPEG standard quantization matrix
  • Lossless compression preserving all DCT coefficients
  • Block-wise DCT/IDCT operations (8×8 blocks)
  • Compression ratio analysis and file size comparison
• MNIST Digit Classification: CNN implementations for handwritten digit recognition
  • Basic 3-layer CNN architecture
  • Enhanced CNN with BatchNormalization, Dropout, and L2 regularization
  • Data augmentation (rotation, shifts, zoom)
  • Learning rate scheduling and early stopping
• CIFAR-10 Classification: Color image classification with CNN
  • 10-class object recognition on 32×32 color images
  • Similar architecture adapted for RGB inputs
• Model Evaluation: Comprehensive performance metrics
  • Classification reports with precision, recall, F1-score
  • Confusion matrices with heatmap visualization
  • ROC curves and AUC scores

Notebook: Segmentation, Detection & Classification

• File: segmentation_detection_classification.ipynb
• Image Segmentation: Multiple segmentation approaches
  • Edge-based segmentation using Canny edge detection
  • Region-based segmentation with thresholding techniques
  • Visualization with matplotlib for result comparison
• Hough Transform: Line detection and feature extraction
  • Probabilistic Hough Line Transform for straight line detection
  • Configurable parameters for line detection sensitivity
  • Visual overlay of detected lines on original images
• Object Detection: State-of-the-art detection models
  • YOLOv8: Real-time object detection with ultralytics framework
  • Faster R-CNN: Region-based detection with ResNet50-FPN backbone
  • Bounding box visualization with confidence scores
  • Pre-trained models on COCO dataset for 80+ object classes
• Deep Learning Classification: Multi-dataset CNN training
  • Fashion-MNIST: Clothing classification (10 classes, 28×28 grayscale)
  • CIFAR-100: Fine-grained object classification (100 classes, 32×32 RGB)
  • Custom CNN architectures with Conv2D, MaxPooling, and Dense layers
  • 5-epoch training with validation accuracy tracking
  • Classification reports with precision, recall, and F1-scores
• Integrated Pipeline: End-to-end processing workflow combining segmentation, detection, and classification
• Dual Environment Support: Compatible with both local (matplotlib) and Google Colab (cv2_imshow) environments

Notebook: Blob Detection, Image Enhancement & Classification

• File: blob_detection_image_enhancement_classification.ipynb
• Blob Detection Algorithms: Implementation of three advanced blob detection methods
  • LoG (Laplacian of Gaussian): Scale-space blob detection with adjustable sigma parameters
  • DoG (Difference of Gaussian): Efficient approximation of LoG for faster computation
  • DoH (Determinant of Hessian): Hessian matrix-based blob detection for feature localization
  • Purple region extraction using HSV color masking
  • Morphological preprocessing pipeline (erosion, dilation, opening, closing, area operations)
  • Handles RGBA images with alpha channel conversion
  • Red circle overlay visualization of detected blobs
• Image Enhancement Pipeline: Eight comprehensive image processing techniques
  • Brightness & Contrast adjustment with alpha/beta parameters
  • Image sharpening using custom convolution kernels
  • Denoising with Non-Local Means algorithm
  • Color enhancement using PIL ImageEnhance
  • Image resizing with interpolation
  • Inverse transform (bitwise NOT operation)
  • Histogram equalization (grayscale and color via YCrCb)
  • LAB color space-based color correction
  • Grid visualization (3×3 layout) of all enhancement results
• Transfer Learning Classification: CIFAR-100 fine-grained classification
  • AlexNet: Pre-trained on ImageNet, fine-tuned for 100 classes
  • VGG16: Deep architecture with 16 layers, adapted for CIFAR-100
  • Modified final classifier layers for 100-class output
  • SGD optimizer with momentum (lr=0.0001, momentum=0.9)
  • Cross-entropy loss function
  • Training loop with tqdm progress bars
  • Batch size optimized for memory efficiency (batch_size=16)
  • Automatic device selection (CUDA/MPS/CPU)
  • ImageNet normalization for transfer learning compatibility
• Model Evaluation: Comprehensive accuracy metrics on CIFAR-100 test set
• Multi-Image Processing: Batch processing across multiple test images (p1.jpg, p2.jpg, p3.png, p4.png, p5.jpg)

Notebook: SIFT, ORB, Watershed, ResNet & Few-Shot Learning

• File: sift_orb_watershed_resnet_few_shot_learning.ipynb
• Feature Detection & Description: Classical computer vision feature extraction
  • SIFT (Scale-Invariant Feature Transform): Keypoint detection and descriptor computation
  • ORB (Oriented FAST and Rotated BRIEF): Fast binary descriptor for real-time applications
  • Rich keypoint visualization with OpenCV drawing flags
  • Multiple test images for robustness evaluation
• Feature Matching: Correspondence finding between image pairs
  • BFMatcher (Brute-Force Matcher): Exhaustive descriptor matching
  • Hamming distance for ORB binary descriptors
  • L2 distance for SIFT float descriptors
  • Cross-check validation for bidirectional matching
  • Top-50 matches visualization with distance-based sorting
• Watershed Segmentation: Marker-based region segmentation
  • Binary thresholding (THRESH_BINARY_INV) for preprocessing
  • Custom marker seeds for background/foreground separation
  • Watershed algorithm with boundary highlighting (red contours)
  • Contour detection with RETR_EXTERNAL mode
  • Multi-stage visualization (original → watershed → contours)
• Deep Learning Classification: CIFAR-100 fine-grained recognition
  • ResNet-18: Residual network with 18 layers (ImageNet pre-trained)
  • ResNet-34: Deeper 34-layer residual architecture
  • Transfer learning with modified final FC layer (100 classes)
  • SGD optimizer with momentum (lr=0.0001, momentum=0.9)
  • Cross-entropy loss with 5-epoch training
  • ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
  • Training progress tracking with tqdm progress bars
  • GPU/CPU automatic device selection
• Few-Shot Learning: Meta-learning for low-data scenarios
  • Prototypical Networks: Metric learning with prototype computation
    • Simple FC encoder (784→256→64 dimensions)
    • Episode-based sampling (5-way, 5-shot setup)
    • Euclidean distance metric in embedding space
  • MNIST Dataset with 80/20 train/test split
  • Elastic deformation augmentation for data diversity
• One-Shot Learning: Siamese network for similarity learning
  • Siamese CNN: Twin network architecture for pair comparison
    • Convolutional feature extractor (Conv2D→ReLU→MaxPool2D×2)
    • Fully connected similarity head (128×4×4→256→128)
    • Pairwise distance computation with L2 norm
  • Proper tensor flattening for CNN-to-FC transition
• Elastic Transform Augmentation: Advanced data augmentation
  • Corrected Implementation: Gaussian-smoothed displacement fields
  • Alpha scaling for deformation magnitude control
  • Sigma parameter for smoothness adjustment
  • scipy.ndimage.gaussian_filter for proper elastic deformation
  • cv2.remap with bilinear interpolation and reflection borders
  • Side-by-side visualization of original vs. deformed images
• Model Evaluation Metrics: Comprehensive performance analysis
  • Accuracy, Precision, Recall, F1-score from sklearn.metrics
  • Weighted averaging for multi-class scenarios
  • Custom Siamese evaluation with pairwise similarity threshold
  • Confusion matrix support for detailed error analysis
• Dual Mode Support: Compatible with local and Colab environments

Notebook: Stitching, Denoising, GANs & Segmentation Playground

• File: Stitching_Denoising_GAN_SegmentationPlayground.ipynb
• Image Stitching: Simple and panorama modes using OpenCV Stitcher; ORB-based matching with essential matrix pose recovery; visualization of keypoints and matches.
• Inpainting: Mask-based OpenCV inpainting helper for quick cleanup of noisy regions.
• Denoising Autoencoders (MNIST): Two variants (basic and improved with BN/Dropout) for noise+blur restoration; PSNR/SSIM metrics and pixel-wise accuracy.
• GANs: MNIST MLP GAN and CIFAR-10 DCGAN with loss plots and image grid visualization utilities.
• Transfer Learning: MobileNet V1/V2/V3 comparison for dog-breed classification (timm/models), with frozen backbone and classifier fine-tuning.
• Utilities: Notebook metadata cleaning snippets (nbformat) to keep GitHub rendering healthy; Colab upload helpers for stitching inputs.

Notebook: Image Denoising & Video Action Pipeline

• File: Image_Denoising_Video_Action_Pipeline.ipynb
• Task 1: Image Denoising Comparison
  • Median Filter Denoising: Channel-wise morphological filtering with disk-shaped structuring elements (size=3), RGB channel processing, PSNR/SSIM/MSE metrics, matplotlib visualization
  • Wavelet Denoising: BayesShrink soft thresholding in wavelet domain, adaptive sigma rescaling, per-channel decomposition/reconstruction, quality metrics comparison
  • Deep Learning Denoising: Noise2Void-inspired U-Net (Conv2D→MaxPooling→UpSampling), self-supervised patch training (64×64), center pixel masking strategy, 10-epoch training
  • Multi-Method Comparison: Quantitative metrics (MSE, PSNR, SSIM) and visual side-by-side evaluation
• Task 2: Video Processing & Action Recognition
  • Frame Extraction: Configurable interval sampling, sequential frame storage, progress tracking
  • Frame Visualization: 2×5 grid display with BGR→RGB conversion
  • Video Operations: Adaptive thresholding (GAUSSIAN_C), Gaussian blur (5×5), Canny edges (100/200), bitwise NOT inversion
  • Frame Collage: Grid-based spatial sampling, adaptive sizing, half-resolution assembly
  • UCF101 Dataset: 5-class subset (Basketball, Biking, PlayingGuitar, Typing, JumpRope), 10 videos/class, 224×224 frames, 16-frame sequences
  • 3D CNN Classification: 3-layer Conv3D (64→128→256 filters), batch normalization, 2×2×2 max pooling, 512-unit FC + 50% dropout, L2 regularization
  • Training Pipeline: Data augmentation (±30° rotation/shifts/shear/zoom, horizontal flip), class weight balancing, Adam optimizer (lr=0.0001), 100 epochs, ModelCheckpoint callback
  • Evaluation: Test accuracy, per-class precision/recall/F1, confusion matrix, training/validation curves
• Dependencies: PyWavelets, TensorFlow/Keras, scikit-learn, OpenCV, matplotlib
• Colab Ready: Full environment setup with pip installations