← Back to Publications List

MS-YOLOv12: An Enhanced Real-Time Recyclable Waste Detection and Classification Framework

Students & Supervisors

Student Authors
Mehedi Hasan Shuvo
Bachelor of Science in Computer Science & Engineering, FST
Mahbubul Islam
Bachelor of Science in Computer Science & Engineering, FST
Supervisors
Kamruddin Nur
Professor, Faculty, FST
Rifat Tasnim Anannya
Assistant Professor, Faculty, FST
Aminun Nahar
Assistant Professor, Faculty, FST

Abstract

Real-time waste classification and detection are key to utilizing smart recycling systems and minimizing manual labor in waste management processes. Manual sorting is time-consuming, inaccurate, and unreliable. The proposed MSYOLOv12 is a deep learning model that is designed to automatically detect and classify four types of recyclable items: can, glass, plastic, and paperpack. In our proposed model, standard convolution (Conv) is replaced by GhostConv (Ghost Convolution) in layer 7 of the backbone and Depthwise Convolution (DWConv) in layer 18 of the neck to reduce redundancy and computational cost while enhancing spatial-channel feature fusion. The model is trained on the Recycle Computer Vision Project dataset, consisting of 3953 images, using the Roboflow platform with standardized resolution and orientation. The proposed MSYOLOv12 model achieves 92.1% precision, 91.6% recall, 92.0% F1-score, 96.6% mAP@50, and 93.6% mAP@50-95. Moreover, the model reduces the number of parameters by 12.1%, decreases inference time by 5.1%, and improves processing speed by 5.5% compared to the baseline YOLOv12, making it suitable for realtime applications.

Keywords

Machine learning deep learning YOLOv12 artificial intelligence machine vision

Publication Details

  • Type of Publication:
  • Conference Name: International Conference of Frontiers of Engineering and Emerging Technologies (FET’26)
  • Date of Conference: 22/04/2026 - 22/04/2026
  • Venue: University of Bahrain , Kingdom of Bahrain (online)
  • Organizer: University of Bahrain
View on Publisher