Inorganic Trash Sorting Mobile Application using MobileNetV3 Method

Abstract

This study aims to develop a mobile application for inorganic waste sorting capable of detecting seven object classes (plastic, glass, metal, electronics, paper, fabric, and cardboard) using the MobileNetV3 Deep Learning architecture. Based on comparative testing results, the MobileNetV3 method proved to be more effective than EfficientDet-d0, characterized by consistent loss reduction, better computational time efficiency, and optimal generalization capabilities without overfitting. Dataset quality was found to be a fundamental factor in model success, where the use of a standardized and noise-free dataset yielded superior performance with a detection accuracy reaching 97.14% and an Overall mAP of 94.13%. In-depth analysis showed that while detection performance was very stable for fabric and electronics categories, accuracy for plastic, glass, and metal categories tended to decrease when high Intersection over Union (IoU) precision was required. Operationally, the system achieved a 100% success rate within the distance range of 30 cm to 150 cm, with the best time efficiency observed between 60 cm and 90 cm. Furthermore, hardware testing concluded that high specifications only provided significant advantages in long-distance detection (>150 cm), whereas for the ideal operational range, devices with standard specifications proved to be highly sufficient and efficient for system implementation