Deep Learning / Computer Vision

GreenSort – Automated Deep Learning Waste Classifier

Deep convolutional neural network classifying municipal waste into organic and recyclable streams with 92.04% test accuracy.

Hosted on Render free tier (may take ~20s to wake up on initial request).

Problem Statement

Improper municipal waste segregation causes severe landfill overflow and contamination of recyclable materials. Automated, low-latency visual sorting is vital to relieve human sorting burdens and enhance recycling throughput.

Methodology & Implementation

Built and trained a 4-stage convolutional neural network (CNN) in TensorFlow/Keras on a curated dataset of 25,077 real-world images. Replaced heavy dense layers with GlobalAveragePooling2D, reducing model parameter count by 99.3% while maintaining high discriminatory feature representation. Deployed with an asynchronous FastAPI inference service featuring image preprocessing, batch predictions, and automated Swagger OpenAPI documentation.

Technical Architecture Highlights

  • 4-stage convolutional feature extraction hierarchy with batch normalization and dropout regularization
  • GlobalAveragePooling2D topology slashing total model parameters by 99.3% to enable lightweight edge inference
  • Asynchronous FastAPI REST endpoints processing image payloads with sub-50ms inference latency
  • Interactive client dashboard supporting real-time single-item uploads and multi-file batch image classification

Results & Evaluation

01.92.04% test classification accuracy across unseen evaluation sets
02.Trained and validated on 25,077 diverse real-world waste images
03.Sub-50ms CPU inference latency per image on containerized deployments
04.Production API and web interface deployed live on Render

Technologies & Tools

#Python#TensorFlow#Keras#FastAPI#NumPy#OpenCV#Render#Docker