A comprehensive collection of computer vision projects covering semantic segmentation, object detection, image classification, and deep learning techniques.
This repository contains practical implementations and projects completed during an intensive computer vision course at AUCA. Each project demonstrates state-of-the-art computer vision techniques using PyTorch, with a focus on real-world applications.
Semantic segmentation of power line insulators using U-Net
- Task: Segment insulators on aerial power line photos captured by UAVs
- Architecture: U-Net with ResNet34 encoder
- Performance: 0.9895 Dice coefficient
- Techniques: Transfer Learning, Data Augmentation, Test-Time Augmentation
- Framework: PyTorch, Albumentations
Multi-class image classification for grain type identification
- Task: Classify grain images into 4 categories (barley, flax, oats, wheat)
- Architecture: EfficientNetV2, ConvNeXt (ensemble)
- Techniques: 5-fold Cross-Validation, Test-Time Augmentation, Mixed Precision Training
- Framework: PyTorch, timm library
- Key Features: Ensemble methods, learning rate scheduling, data augmentation
Comprehensive notebooks covering the full Deep Learning and Computer Vision curriculum:
- PyTorch Fundamentals - Introduction to PyTorch and tensor operations
- Gradient Descent & Optimization - Optimization algorithms and convergence
- Neural Network Classification - Building classifiers from scratch
- Advanced Classification Techniques - Ensemble methods and regularization
- CNN From Scratch - Implementing convolutional layers and forward pass
- Fully Connected vs CNN - Comparing dense and convolutional architectures
- AlexNet Architecture - Classic CNN implementation and analysis
- ResNet & VGG Architectures - Deep residual networks and VGG
- Object Detection Methods - Detection techniques and frameworks
- DL Optimization & Regularization - Advanced training techniques
- Semantic Segmentation: U-Net & SegNet - Pixel-wise classification with encoder-decoder architectures
computer_vision_course/
├── README.md # This file
├── Murat_Raimbekov_CV.pdf # Resume/CV
├── requirements.txt # Common dependencies
│
├── 01_PyTorch_Fundamentals.ipynb # Course notebooks
├── 02_Gradient_Descent_Optimization.ipynb
├── 03_Neural_Network_Classification.ipynb
├── 04_Advanced_Classification_Techniques.ipynb
├── 05_CNN_From_Scratch.ipynb
├── 06_Fully_Connected_vs_CNN.ipynb
├── 07_AlexNet_Architecture.ipynb
├── 08_ResNet_VGG_Architectures.ipynb
├── 09_Object_Detection_Methods.ipynb
├── 10_DL_Optimization_Regularization.ipynb
├── 11_Semantic_Segmentation_UNet_SegNet.ipynb
│
├── 01_insulator_segmentation/ # Projects
│ ├── insulator_segmentation.ipynb
│ ├── requirements.txt
│ ├── README.md
│ └── samples/
│
└── 02_grain_classification/
└── grain_classification.ipynb
- PyTorch - Primary framework for all models
- TorchVision - Pre-trained models and transforms
- timm - PyTorch Image Models library
- scikit-learn - ML utilities and metrics
- OpenCV - Image processing and manipulation
- Albumentations - Advanced data augmentation
- scikit-image - Image processing algorithms
- NumPy - Numerical computations
- Pandas - Data manipulation
- Matplotlib - Plotting and visualization
- Seaborn - Statistical visualizations
- Jupyter Notebook - Interactive development
- Git/GitHub - Version control
- CUDA - GPU acceleration
- ✅ Semantic Segmentation (U-Net, SegNet)
- ✅ Image Classification (CNN, EfficientNet, ConvNeXt)
- ✅ Object Detection (YOLO - see hackathon project)
- ✅ Transfer Learning
- ✅ Data Augmentation & Preprocessing
- ✅ Model Training & Optimization
- ✅ Cross-Validation & Ensemble Methods
- ✅ Test-Time Augmentation
- ✅ Mixed Precision Training
- ✅ Learning Rate Scheduling
- ✅ Loss Function Design (BCE, Dice, Focal)
- ✅ Performance Evaluation & Metrics
- ✅ PyTorch Model Architecture Design
- ✅ Custom Dataset & DataLoader Implementation
- ✅ Training Pipeline Development
- ✅ Ablation Studies & Experimentation
- ✅ Result Visualization & Analysis
Yandex Hackathon Project - December 2024
Automated road defect detection system using YOLOv8 for real-time object detection on highway images.
- Achievement: 85%+ mAP on test set
- Technologies: YOLOv8, PyTorch, OpenCV, Albumentations
- Repository: hackathon-urban-tech
pip install -r requirements.txtEach project has its own dependencies. Navigate to the specific project directory:
cd 01_insulator_segmentation
pip install -r requirements.txtFor CUDA/GPU acceleration:
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118jupyter notebook 01_PyTorch_Fundamentals.ipynbcd 01_insulator_segmentation
jupyter notebook insulator_segmentation.ipynb| Project | Task | Best Model | Metric | Score |
|---|---|---|---|---|
| Insulator Segmentation | Semantic Segmentation | U-Net + ResNet34 | Dice Coefficient | 0.9895 |
| Medical Image Segmentation | Skin Lesion Detection | SegNet + BCE | IoU | 0.654 |
| Grain Classification | Multi-class Classification | Ensemble (EfficientNetV2 + ConvNeXt) | F1-Score | High |
| Road Defects Detection | Object Detection | YOLOv8 | mAP | 85%+ |
Murat Raimbekov Data Science & Computer Vision Intern
This repository is for educational and portfolio purposes.
- American University of Central Asia (AUCA) - Neural Networks and Deep Learning Course
- Course instructor
- Open-source community (PyTorch, OpenCV, Albumentations)