The Deep Learning Image Classifier is an interactive educational platform designed to demystify the inner workings of Convolutional Neural Networks (CNN) and Transfer Learning. Built with Streamlit and TensorFlow, the application provides real-time visualizations of image preprocessing, RGB channel analysis, convolution operations, and intermediate feature maps, allowing users to watch as the model extracts features and reaches a classification decision.
Tech Stack
Python
TensorFlow
Keras
Streamlit
Plotly
NumPy
SciPy
MobileNetV2
Pillow
Tools Used
Jupyter Notebook
Google Colab
Streamlit Cloud
VS Code
Key Features
Interactive Analytics
▸RGB Matrix Analysis: Dynamic heatmaps and histograms for detailed color channel decomposition.
▸Feature Map Extraction: Live visualization of intermediate layer outputs to understand what the model "sees".
▸Prediction Confidence: Interactive probability distributions and confidence meters for classification results.
Architecture Deep-Dive
▸Custom CNN Workflow: Step-by-step breakdown of a 3-layer convolutional network for hand gesture recognition.
▸Conv Operations: Real-time demonstration of filters extracting edges, textures, and patterns.
▸MobileNetV2 Integration: Implementation of pre-trained ImageNet models for high-accuracy Cheetah vs Hyena classification.
Model Learning
▸Transfer Learning Mechanics: Visual explanation of frozen base models and custom classifier fine-tuning.
▸Hyperparameter Insights: Interactive Sliders to adjust visualization parameters and model thresholds.
▸Performance Reporting: Detailed accuracy and loss metrics comparison between custom and pre-trained architectures.
Educational UX
▸Progressive Visualization: Managed learning path from raw image input to final classification probability.
▸Interactive Documentation: Integrated explanations of convolution, pooling, and activation functions.
▸Dual-Model Comparison: Parallel demonstration of standard CNN vs state-of-the-art Transfer Learning (MobileNetV2).
Constructed sub-models using the Keras Functional API to extract output tensors from specific intermediate layers without interrupting the main inference pipeline.
Challenge
High-Resolution Heatmap Lag
Solution
Optimized RGB intensity distributions using Plotly's WebGL-accelerated rendering and data downsampling for ultra-smooth interactive charts.
Challenge
Limited Computation on Edge
Solution
Utilized depthwise separable convolutions from MobileNetV2 to achieve 96%+ accuracy while maintaining low-latency inference on standard CPU-based instances.