Development of YOLOv8n-cls-Based Intelligent Waste Classification System for Smart Recycling Applications

Image-based classification offers a potential approach to improving waste segregation; however, model selection requires consideration of both classification accuracy and computational efficiency. This study assessed YOLOv8n-cls for classifying household waste into six categories and benchmarked its performance against MobileNetV2, ResNet50, and EfficientNetB0 using an identical, hash-verified test set. Following image quality checks and duplicate removal, 2,524 images were included, comprising cardboard (403), glass (501), metal (409), paper (594), plastic (480), and trash (137). The dataset was divided into training (70%), validation (15%), and testing (15%) subsets using a fixed random seed of 42. Each model employed pretrained weights and an input resolution of 224 × 224 pixels. Among the evaluated architectures, YOLOv8n-cls produced the highest test accuracy (92.15%) and macro-F1 score (0.910), outperforming ResNet50 (89.01%; 0.869), EfficientNetB0 (88.22%; 0.858), and MobileNetV2 (84.29%; 0.810) in observed performance. Its macro one-vs-rest ROC-AUC reached 0.992. However, Holm-adjusted McNemar tests identified a statistically significant difference only between YOLOv8n-cls and MobileNetV2 (p = 0.0054); differences relative to ResNet50 (p = 0.3636) and EfficientNetB0 (p = 0.3457) were not statistically significant. Testing on a Tesla T4 GPU yielded a median single-image inference latency of 6.38 ms, with 1.446 million parameters and a checkpoint size of 2.98 MB. Despite achieving the best observed predictive performance, YOLOv8n-cls did not demonstrate statistically significant superiority over the two stronger CNN benchmarks. The findings should also be interpreted in light of the dataset's limited size, class imbalance, public-source origin, and the absence of field validation. Accordingly, the model is best regarded as a research prototype rather than an operationally validated waste-sorting solution.

Authors

Institutions

Publication Details

Journal
IJACI International Journal of Advanced Computing and Informatics
Published
2026-11-20
DOI
https://doi.org/10.71129/ijaci.v3i1.pp68-85
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Development of YOLOv8n-cls-Based Intelligent Waste Classification System for Smart Recycling Applications

David Juacian Alcarde, James Costiniano Alcarde, Jancen Cesar D. Tabag
4 links
IJACI International Journal of Advanced Computing and Informatics
Advanced Neural Network Applications
article

Development of YOLOv8n-cls-Based Intelligent Waste Classification System for Smart Recycling Applications

David Juacian Alcarde, James Costiniano Alcarde, Jancen Cesar D. Tabag
article en

Abstract

Image-based classification offers a potential approach to improving waste segregation; however, model selection requires consideration of both classification accuracy and computational efficiency. This study assessed YOLOv8n-cls for classifying household waste into six categories and benchmarked its performance against MobileNetV2, ResNet50, and EfficientNetB0 using an identical, hash-verified test set. Following image quality checks and duplicate removal, 2,524 images were included, comprising cardboard (403), glass (501), metal (409), paper (594), plastic (480), and trash (137). The dataset was divided into training (70%), validation (15%), and testing (15%) subsets using a fixed random seed of 42. Each model employed pretrained weights and an input resolution of 224 × 224 pixels. Among the evaluated architectures, YOLOv8n-cls produced the highest test accuracy (92.15%) and macro-F1 score (0.910), outperforming ResNet50 (89.01%; 0.869), EfficientNetB0 (88.22%; 0.858), and MobileNetV2 (84.29%; 0.810) in observed performance. Its macro one-vs-rest ROC-AUC reached 0.992. However, Holm-adjusted McNemar tests identified a statistically significant difference only between YOLOv8n-cls and MobileNetV2 (p = 0.0054); differences relative to ResNet50 (p = 0.3636) and EfficientNetB0 (p = 0.3457) were not statistically significant. Testing on a Tesla T4 GPU yielded a median single-image inference latency of 6.38 ms, with 1.446 million parameters and a checkpoint size of 2.98 MB. Despite achieving the best observed predictive performance, YOLOv8n-cls did not demonstrate statistically significant superiority over the two stronger CNN benchmarks. The findings should also be interpreted in light of the dataset's limited size, class imbalance, public-source origin, and the absence of field validation. Accordingly, the model is best regarded as a research prototype rather than an operationally validated waste-sorting solution.

IJACI International Journal of Advanced Computing and InformaticsVol. 3(1)
Department of Science and Technology (PH), Industrial Technology Development Institute (PH)
Responsible consumption and production
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon