CNN Implementation in Progressive Web App for Automatic Garbage Classification using TensorFlow.js

Abstract

The substantial and continuously increasing volume of global waste has become a critical environmental challenge, exacerbating the inherent inefficiency of conventional manual sorting techniques. This research addresses this problem by developing and evaluating an automated waste classification system using Convolutional Neural Networks (CNN), specifically the VGG16 architecture, integrated into a Progressive Web App (PWA) to enhance accessibility and sorting efficiency. Our primary goal is to deliver an intelligent, lightweight, and cross-platform solution capable of performing client-side inference on diverse devices. The VGG16 model was retrained using transfer learning on a validated public dataset of 10,365 images, comprising two classes (organic and inorganic waste). The trained model was converted to a browser-compatible format, TensorFlow.js, and deployed within the PWA framework which utilizes Service Workers for offline capabilities. Despite the significant challenge posed by the VGG16 model's large size, the system successfully performed client-side inference by prioritizing GPU acceleration and achieved 0.94 overall accuracy on the test dataset². This result, supported by high F1-scores for both waste categories, confirms that deploying high-accuracy CNN models at the edge using PWA and TensorFlow.js is a feasible and promising strategy for practical, technology-based waste management and environmental education.