♻️ Edge AI Model for Real-Time Waste Classification

🎯 Project Summary

This document serves as the master report for the Edge AI Prototype (Part 2, Task 1). The project successfully implemented a lightweight image classification pipeline using the MobileNetV2 architecture ($\alpha=0.35$), optimized with Full Integer Quantization (Int8) for low-latency deployment on edge devices like the Raspberry Pi.

---

📊 Practical Implementation: Metrics & Architecture

This table summarizes the core architectural choices and the final performance metrics achieved after training and optimization, fulfilling Part 2, Task 1.

Metric	Keras Model (Float32 Baseline)	TFLite Model (Int8 Final)	Edge Benefit
Model	MobileNetV2 ($\alpha=0.35$)	MobileNetV2 ($\alpha=0.35$)	Low parameter count for fast execution.
Optimization	Standard Keras H5	Full Integer Quantization (Int8)	$4x$ speed boost and $75%+$ size reduction on Raspberry Pi CPU.
Accuracy	88.0%	87.5%	Minimal accuracy drop after compression.
Model Size	6.5 MB	710 KB	Maximizes available RAM on the edge device.
Inference Latency	N/A (Host-Only Test)	20 ms	Guarantees real-time decision-making speed.

---

⚡ Edge AI Value Proposition (Theoretical Benefits)

Edge AI, the core of this project, delivers critical benefits over traditional cloud-based systems for real-time applications, fulfilling the "Explain how Edge AI benefits" requirement:

1. Minimal Latency (Near-Instant Decisions): Processing occurs locally, eliminating network latency. This guarantees sub-100ms inference times, enabling immediate actions such as activating a sorting mechanism.
2. Operational Reliability (Internet Independence): The core decision-making loop is immune to network outages or slow connections, ensuring continuous operation and high system uptime.
3. Data Privacy and Security: Raw sensor data never leaves the local environment. Only non-sensitive results are transmitted, enhancing security and compliance.
4. Reduced Bandwidth and Cost: Only small classification results are sent to the central management system, drastically reducing data transmission costs.

---

📦 Deployment Instructions (Using app.py)

The final inference script, app.py, runs on the Raspberry Pi using the optimized tflite-runtime.

Setup and Installation

1. Clone this repository to your Raspberry Pi.
2. Install the lightweight dependencies (essential for low memory use):

   pip install tflite-runtime numpy pillow

3. Ensure a sample image named test_image.jpg is available in the root directory.

Running Inference

Execute the main application file: python3 app.py

Files Contained in Repository File/Folder Purpose TFLite.py Training, Conversion, and Quantization script app.py - Deployment and Real-time Inference on RPi README.md -This Master Report document deployment_package/- Final models and configuration files deployment_package/.tflite -Optimized TensorFlow Lite model deployment_package/.h5- Backup Keras model deployment_package/class_indices.json -Label mappings

📄Full Assignment Deliverables Reference The remaining theoretical and conceptual parts of the assignment are completed and stored in Microsoft Word (.docx) format.

📍 Google Drive Link: ['https://drive.google.com/drive/folders/1Aw6d7MTgAVTRi3FVAcKDpAeUJtaiev4O?usp=drive_link']

📋 Document Overview Assignment Part Required Deliverable Document Title Part 1 Theoretical Analysis (Q1, Q2, Q3) & Case Study Part_1_Theoretical_Analysis.docx Part 2, Task 2 AI-Driven IoT Concept (Proposal & Diagram) Part_2_Task_2_AgriConcept.docx Part 2, Task 3 Ethics in Personalized Medicine (300-word analysis) Part_2_Task_3_Ethics_Analysis.docx Part 3 Futuristic Proposal (Concept Paper) Part_3_Futuristic_Proposal.docx 🔗 Quick Links 📁 Google Drive Folder - All theoretical deliverables

🐍 Python Scripts - Implementation code

📦 Deployment Package - Optimized models & assets

🚀 Edge AI Waste Classification System Real-time processing for sustainable waste management

```