Research on marine litter detection based on CNN-Transformer heterogeneous parallelism

Kui Chen , Chenglin Luo, Yuhong Tang

Aiming at strong background interference and low detection accuracy of small/deformed targets in marine debris detection, this paper develops a high-precision lightweight intelligent detection and recognition system. A multi-scenario dataset is built and data augmentation is used to tackle sample scarcity and domain shift; a CNN-Transformer heterogeneous parallel model YOLO-Trans is designed on the YOLOv8 baseline for local-to-global feature extraction and accurate detection of small/deformed targets, and a visual detection system is developed with PyQt5. Experiments show the model surpasses the original YOLOv8s in all metrics on the self-built dataset, and ablation experiments validate the improved modules’ effectiveness, offering technical support for large-scale intelligent marine debris monitoring.