Enhancing Waste Detection Accuracy with CNN-based YOLO Model

Indonesia generated over 60 million tons of waste in 2024, with organic ($41.6 \%$) and plastic ($18.7 \%$) waste being the prevalent types. Low accuracy of existing automated detection systems is one of the major issues in waste management, especially when objects are small, backgrounds are cluttered, and illumination varies. This paper proposes a conceptual framework to combine Convolutional Neural Networks (CNNs) with the recently created YOLOv12 model to improve multi-class waste detection accuracy. An examination of recent open-access research confirms that such upgradations as multi-resolution methods (Ren et al., 2024), lightweight structures (Liu et al., 2024; Li et al., 2024), and advanced attention mechanisms (Yang et al., 2024) can boost mean Average Precision (mAP) by $5-12 \%$ compared to the earlier YOLO versions. In addition, CNNs have proven to be highly resilient in their ability to extract strong features even from poor data conditions (Yudita et al., 2021), which further makes them solid backbones for advanced object detection. This study provides a conceptual framework which has yet to be tested and proven empirically, but it constructs a theoretical foundation upon which the CNN-based YOLOv12 model may be studied and deployed into real-world waste detection systems in the future. This conceptual study places CNN-based YOLOv12 as a potential solution towards real-time waste detection and aims to contribute towards constructing smart waste management systems in Indonesia.