Smart Garbage Classification System Based on MobileNetV2: Real-Time Waste Recognition Without Training Data

This project is Pemilahan-sampah-ai (Indonesian AI Garbage Classification) on GitHub, developed by rezaachmad-sketch and released on June 3, 2026, under the Educational License. The core highlight is using the pre-trained MobileNetV2 model and keyword mapping technology to recognize organic and non-organic garbage in real time without custom training datasets, demonstrating the practical value of transfer learning in the environmental protection field. Original link: https://github.com/rezaachmad-sketch/Pemilahan-sampah-ai.

Against the backdrop of environmental protection and smart city construction, garbage classification has become a global focus. However, traditional AI classification systems usually require large amounts of labeled data and a long training process. This project provides an innovative and practical solution: through pre-trained MobileNetV2 model + keyword mapping technology, it bypasses the thresholds of data collection and training, enabling rapid deployment of a usable real-time garbage classification system.

Reasons why the project does not require training:

1. Leverage MobileNetV2's generalization ability on the large-scale ImageNet dataset (capable of recognizing 1000 common objects)
2. Establish mapping rules from object names to organic/non-organic categories
3. Directly use the pre-trained model for real-time inference

Mapping rule examples:

Category	Recognizable Object Examples
Organic (Organik)	banana, apple, corn, leaf, food
Non-Organic (Non-Organik)	bottle, can, plastic, paper, cardboard

This design uses the visual features of the pre-trained model and completes classification through a simple rule engine, eliminating the need to train a dedicated classifier from scratch.

Technology Stack Composition:

Component	Purpose	Version/Description
Python	Development language	3.13
TensorFlow/Keras	Deep learning framework	Run MobileNetV2
OpenCV	Computer vision	Real-time image processing
MobileNetV2	Pre-trained model	Lightweight CNN architecture

System Workflow: Load pre-trained model → Activate camera → Capture video frame → Crop central area → Image preprocessing → Model inference → Keyword mapping classification → Display results

Key design decisions: Central area cropping (reduces interference, focuses on targets, lowers computational load); Direct use of pre-trained weights (general features are sufficient to support coarse-grained classification).

MobileNetV2 Advantages:

• Lightweight (about 14MB), suitable for resource-constrained devices
• Fast inference speed, can run in real time on CPU
• Optimized architectures like depthwise separable convolution and inverted residual structure

Project Limitations:

• Coarse classification granularity (only distinguishes between organic and non-organic)
• Dependent on pre-trained model labels (only recognizes relevant objects among the 1000 categories)
• Boundary classification issues exist (e.g., food-stained tissues)
• Recognition accuracy is affected by lighting, angle, and background

Comparison with other models: MobileNetV2 is smaller and faster than ResNet50, making it suitable for real-time camera applications.

Applicable Scenarios: Educational demonstrations, prototype verification, home garbage classification assistants, quick demo system construction for workshops/hackathons

Short-term Improvements: Expand keyword mapping table, add confidence threshold, multi-frame voting to improve stability, user feedback mechanism

Long-term Evolution: Transfer learning fine-tuning (using a small number of local garbage images), multi-category classification (recyclable/hazardous, etc.), object detection upgrade (supports multi-item recognition), edge deployment (Raspberry Pi/AI chips).

Educational Value: A practical example of transfer learning, demonstrating the power of pre-trained models and rapid prototyping methods, inspiring beginners to make good use of open-source resources and prioritize engineering thinking.

Environmental Significance: Helps with correct garbage classification, which can increase recycling rates by 30-50%, reduce landfill volume and processing costs, and lower greenhouse gas emissions.

Summary: The project is small but refined. Its advantages are zero training requirements, mature technology stack, and clear code; limitations are coarse granularity and insufficient robustness. It is suitable for AI beginners to learn or developers to quickly validate concepts.