The accuracy of underwater garbage detection and identification plays a very important role in improving the garbage cleaning work carried out by underwater robots. Based on this, this paper proposes an improved underwater trash detection network model based on YOLOv5s. In order to improve the recognition performance of underwater garbage images, this paper also proposes a weighted fusion-based underwater image enhancement algorithm, which fuses the CLAHE algorithm and Retinex algorithm on the basis of weighted logarithmic transformation and adaptive Gamma correction, so as to improve the quality of underwater garbage images. For underwater garbage detection, GhostNet is introduced to improve the backbone network of YOLOv5s to enhance the feature extraction capability, and combined with the ECA attention mechanism and CARAFE up-sampling mechanism to further realize the model lightweighting and enrich the features and semantic features. The results show that the YOLOv5s-G-E-C model improves the detection average accuracy from 60.92% to 86.19% and the model computation reduces the model computation from 18.42 GLOPs to 15.78 GLOPs compared to the YOLOv5s model.It is feasible to apply the improved YOLOv5s model to underwater garbage detection with better detection performance.