Detection of Sellable Non-Organic Waste Using YOLOv4 Based on Flask Web

Abstract

Non-organic waste management in Islamic boarding school environments requires a sorting system that is fast, accurate, and easy to use, particularly for waste types that have selling value. In the Al-Hasyimiyah area of Nurul Jadid Islamic Boarding School, waste sorting has been implemented to help maintain environmental cleanliness and support the reuse of economically valuable waste. However, the sorting process still faces challenges because some students are not yet able to distinguish types of non-organic waste based on category and selling price. This study aims to develop an image-based non-organic waste detection system to estimate selling prices using the You Only Look Once version 4 (YOLOv4) method implemented in a Flask-based web application. The dataset consisted of 1,170 JPG images and one MP4 video with five waste classes: aluminium, mixed plastic container, plastic bottle, plastic cup, and cardboard. The dataset was divided into 819 training data, 234 validation data, and 117 testing data. The research stages included problem identification, dataset collection, preprocessing, image resizing to 416 × 416 pixels, annotation using LabelImg, data splitting, YOLOv4 training, model testing, selling price estimation, and Flask web implementation. The experimental results showed that the best configuration was obtained using batch 64 and max_batches 2500, achieving an mAP of 90.70%, precision of 79%, recall of 89%, and F1-score of 84%. Testing on 251 objects produced 226 detected objects and 25 undetected objects, resulting in an overall accuracy of 90%. The findings indicate that YOLOv4 can be effectively used as a supporting tool for detecting and estimating the selling price of non-organic waste