End-to-End AI Waste Classification System | Custom CNN & FastAPI Backend

Executive Overview

An end-to-end computer vision solution engineered to automate waste sorting and material classification into five canonical categories: Glass, Metal, Organic, Paper, and Plastic. Designed with production readiness in mind, this project bridges the gap between deep learning model architecture, automated data auditing, and real-time backend API integration.

Key Technical Highlights & Capabilities

Custom CNN Architecture Built from Scratch: Designed a lightweight 4-layer Convolutional Neural Network in TensorFlow/Keras optimized for efficient CPU-bound inference without requiring dedicated GPU resources.

Data Pipeline & Audit Framework: Engineered an automated preprocessing pipeline over 3,900+ images derived from the RealWaste dataset. Implemented SHA-256 cryptographic hashing to eliminate duplicates, corrupt image validation, and balanced class-weight computation.

Production-Ready FastAPI Backend: Built a high-performance RESTful API (/predict and /health endpoints) featuring robust error handling, multipart payload validation, and instant JSON response parsing.

Model Registry & Cloud Hosting: Decoupled model storage by publishing versioned weights to Hugging Face Hub, enabling automated weight retrieval without bloating the core application codebase.

Interactive Web Interface: Developed a clean, responsive HTML/JS client interface allowing users to drag and drop images, stream requests to the backend, and view real-time probability distributions across all classes.

Generalization & Regularization: Incorporated train-time data augmentation (Random Flip, Rotation, Zoom, Contrast) alongside Dropout and Early Stopping callbacks to prevent overfitting.

Tech Stack & Tooling

AI & Computer Vision: Python 3.12, TensorFlow, Keras, OpenCV, NumPy, Pandas.

Backend Development: FastAPI, Uvicorn, REST APIs, Multipart Processing.

Frontend: HTML5, CSS3, JavaScript (Fetch API).

Model Registry & DevOps: Hugging Face Hub, Git, Pytest.

System Architecture Workflow

Input Preprocessing: Standardized 224x224 RGB image rescaling and augmentation.

Feature Extraction: 3x Conv2D blocks (32 → 64 → 128 filters) with ReLU activation and Max-Pooling.

Classification Head: Dropout (0.25) → GlobalAveragePooling2D → Dense (128) → Dropout (0.5) → Softmax (5 Classes).

API Response: Instant prediction payload returning predicted class, confidence percentage, and full probability mapping.

3. Category

Artificial Intelligence & Machine Learning (or Web & Application Development)