Marine debris constitutes a significant global environmental challenge. Autonomous underwater robots offer a potential solution for the removal of deep-sea debris. To enable rapid and accurate detection of such debris， this study presents a lightweight detection model， termed Deep-sea Debris YOLO （Debris-YOLO）， developed using deep learning techniques. An initial deep-sea debris dataset was constructed based on the Deep-sea Debris Database provided by the Global Oceanographic Data Centre （GODAC）. Subsequently， an enhanced BiFPN feature fusion network was employed to reduce model parameters while enhancing background discrimination capabilities. A lightweight detection head was designed to decrease computational complexity， thereby improving the model’s practicality and deployability for deep-sea debris detection. Furthermore， the Wise-DIoU （Wise Distance Intersection over Union） loss function was introduced to mitigate the influence of low-quality samples， enabling more precise localization of deep-sea debris. Data augmentation and adaptive color-balanced underwater restoration were applied to enrich the training dataset. Experimental results demonstrate that the proposed Debris-YOLO model achieves improvements of 1.3% and 1.6% in mAP_0.5 and mAP_0.5∶0.95， respectively， compared to YOLOv8n， while reducing the number of parameters and GFLOPS by 48.2% and 36%， respectively.