<![CDATA[Accurate trash detection in aquatic environments remains a significant challenge for detection models, which exhibit persistent limitations in identifying small and partially submerged objects. Additionally, a notable gap exists in methodologies for fine-tuning the detection model to optimize performance for a specific waterways. To address these limitations, the first objective is to develop a detection model designed to enhance performance on small and partially submerged trash, and the second is to establish a framework for efficiently adapting the model to achieve high accuracy within local waterways. First, the YOLOv11 architecture is enhanced by integrating LCAM and LCBHAM attention mechanisms and pre-trained on various combinations of public datasets to establish a robust, baseline model. For the second objective, this baseline model is adapted using a data-efficient framework. This study process introduces the BojongTrash dataset, captured from a specific waterway, and involves systematically fine-tuning the model on incremental subsets of this data to determine the minimum quantity of images and training epochs required to achieve high accuracy in the target environment. The proposed YOLOv11s-LCA architecture demonstrated a statistically validated improvement over its baseline, increasing the mAP50 score from 0.779 to 0.836 on the FloW-Img dataset with only a 0.1% parameter increase. Furthermore, the research establishes a highly efficient fine-tuning framework, demonstrating peak mAP50 performance of 0.908 that achieved by fine-tuning on 1,000 images for only 3-5 epochs. Therefore, this research validates lightweight attention mechanisms as an efficient strategy for enhancing detection in complex environments and provides a practical framework that enables the rapid deployment of tailored, high-accuracy monitoring systems.]]>
Copyrights © 2026
