Deposits of trash in aquatic environments have devastating effects on marine ecosystems, posing long-term environmental threats. This causes extensive damage to the marine life as well. From the literature, it is identified that small object detection and efficient feature extraction in complex underwater environments still pose challenges. In the proposed work, an improved version of the YOLOv8 underwater trash detection algorithm (YOLO-UTrash) is introduced. The initial step involves utilizing the Density-Based Spatial Clustering of Applications with Noise removal algorithm for real-frame clustering to detect outliers in an unconstrained environment. The PANet layer in the YOLOv8 architecture increases the computational cost by adding an additional bottom-up channel. To address this challenge, an enhanced version of the Bidirectional Feature Pyramid Network (BiFPN) module is incorporated instead of PANet. The eBiFPN’s design is prominent, and its efficiency and ease of integration with other networks enhance feature extraction capabilities. Experimental methodology is tested on the ICRA19 and UIEB, comprising diverse marine pollution scenarios, ensuring the proposed models’ robustness and adaptability. By comparing the proposed work with existing leading models, it is shown that this work is able to outperform by achieving 93% precision, 91% recall, and 94% accuracy, maintaining real-time capabilities essential for timely intervention and remediation efforts.