Abstract: Aiming at the problems of complex seafloor environment where litter targets present multi-scale morphological distribution， inter-class ambiguity， due to high similarity with marine organisms' features， and the resulting issues of insufficient feature extraction capability， poor detection accuracy， and inaccurate localization， a seabed litter target detection algorithm based on the improved YOLOv8 was proposed. First， the ODConv full-dimensional dynamic convolution plus was fused into the C2f of the neck network， to form a new module C2f_ODConv， which enables the model to realize all-round dynamic adjustment of the convolution kernel and more finely adapt to the features of the input data， thus improving the effectiveness of the feature extraction； second， a deformable attention was introduced after the C2f_ODConv， which effectively captured local details in the image and improves the model detection accuracy； finally， UIoU Loss was used instead of CIoU Loss and a linear recession strategy was adopted to localize the target and improve the model generalization ability further accurately. Experiments were conducted on the public dataset TrashCan-Instance， and the experimental results showed that the improved model had a Recall and mAP of 64.4% and 70.4% respectively， which were 4.5 and 2.2 percentage points higher than the baseline model YOLOv8， and also met the underwater spam detection demand.