How mdinadia/vehicle_type_classification_vgg-16_svm works
Project Type and Purpose
• This is a machine‐learning research pipeline reproducing/extending a published paper.
File Structure
• Top level contains:
• VGG_SVM12.ipynb – the Colab notebook orchestrating the entire pipeline
• report.pdf – the final write-up
• README.md – project overview
• No separate “src” or “backend” directories—logic lives inside the notebook.
Extract Key Components from README
• Data source and balancing
• Preprocessing (resize, normalize, split, augment)
• Feature extraction modules (VGG-16, AlexNet, CNN)
• Classification modules (softmax, SVM linear & RBF)
• Evaluation & visualization
Main Components & Their Interactions
a. Data Ingestion & Storage
• External dataset (Kaggle) → Local data store (image folders)
b. Preprocessing
• Resize, normalize, augment → produces train/test image sets
c. Feature Extraction (deep models)
• VGG-16 (no top layers) → feature vectors
• AlexNet → feature vectors
• Custom CNN → softmax features
d. Classifier Training
• Softmax head on VGG/AlexNet/CNN
• SVM module (linear & RBF kernels) using extracted vectors
e. Evaluation & Visualization
• Confusion matrices, histograms, performance plots
f. Reporting
• Notebook outputs → report.pdf
Architectural Patterns & Principles
• Modular pipeline: separate data → preprocessing → feature extraction → classification → evaluation
• Reusability: extracted feature vectors are saved and reused for different classifiers
• Experiment comparability: multiple “branches” (VGG only, AlexNet, CNN, VGG+SVM)
• Orchestration via a single Jupyter/Colab notebook
1. Data Source
2. Data Preprocessing
3. Feature Extractors (branch for each)
4. Classification
• “Softmax Classifier”
• “SVM (Linear & RBF)”
– From AlexNet and CNN: “Softmax Classifier”
5. Evaluation & Visualization
6. Report Generation
Detail Without Over-Segmentation
• Reuse of feature vectors to highlight modularity
• Keep the flow clear and annotate all connectors