Underwater trash detection is considered a key technology for underwater robots to handle underwater trash. However, the complexity and variability of the underwater environment, poor lighting conditions, the loss of detailed information, and the poor performance of low-resolution images due to step-length convolution in traditional CNN(convolutional neural networks) models limit the accuracy and speed of existing models. In order to solve these problems, a novel underwater rubbish detection algorithm named SPDC-YOLOv8 (small proposal detection convolution-YOLOv8) was proposed. The algorithm employed an adaptive spatial decomposition-based CNN module SPD-Conv(space-to-depth convolution) in the backbone network of YOLOv8, replacing the step-length convolution, thus improving the accuracy of the model for low-resolution images and small object detection. Meanwhile, the CARAFE(contentaware reassembly of features) was employed in the up-sampling process of the model, which enhanced the semantic information and feature representation of underwater rubbish, thus improving the robustness of object detection. The experimental results demonstrate that the method proposed achieves 98.6% and 91.2% accuracy on the trash_ICRA19 dataset and TrashCan dataset, respectively, and improves by 0.3% and 0.8% compared to the original YOLOv8 model, with a the computation time is 2.6 ms in both cases. The improved YOLOv8 algorithm proposed is found to be more adaptable to the complex underwater environment.