Abstract: In the process of underwater target detection, the presence of significant noise in underwater images and the mixture of multi-scale targets can easily lead to missed detections and false positives. An improved underwater target detection algorithm called ME-RTDETR（Multi-scale feature Enhancement fusion RTDETR）is proposed, which is based on RTDETR（Deal-Time Detection Transformer） with enhanced multi-scale feature fusion. Firstly, CSPnet is used to replace the original backbone network to enhance the model's feature extraction capabilities. Secondly, the SHViTCGLU module is introduced to replace the residual module in C2f. This module enhances both local and global feature extraction, thereby improving the model's ability to handle noisy images. Thirdly, a multi-scale feature enhancement fusion pyramid structure is designed based on the original CCFM module. The enhanced feature layers and fine-grained feature fusion enable the model to more accurately locate and recognize targets of different scales. Finally, the Focaler-MPDIoU loss function is designed to replace the original GIoU loss function, which allows the model to better learn the position and shape of bounding boxes and accelerates convergence. Experimental results show that the improved model has an accuracy increase of 4.5% compared to the baseline model on the DUO dataset, an average precision value mAP@0.5 increase of 2.1%, and a parameter reduction of 26%. The proposed algorithm has significant improvements in improving average detection accuracy and reducing model parameters, verifying its practical value in underwater target detection tasks.