Abstract: To address the current problems such as low efficiency of manual garbage collection, high labor costs, and reduced accuracy of garbage detection due to the complexity of the underwater environment, an underwater garbage detection algorithm based on improved RT-DETR network is proposed. Firstly, the lightweight network StarNet is adopted to replace the original backbone network to achieve the simplification of the model. Secondly, a new feature pyramid structure is designed, aiming to enhance the feature information of small targets, replacing the traditional method of adding P2 layers. It also integrates the CSPO (CSP-OmniKernel) module and the SPD convolution module to improve the model's extraction of global features and the fusion of multi-scale features. In addition, the WaveletUnPool module and the LDConv module are introduced to reduce the loss of feature information and optimize the upsampling and downsampling operations to further enhance the accuracy of small target detection. Finally, the Focaler-MPDIoU loss function is designed to replace the loss function of the original model, assigning different weights to samples of different difficulties, which optimizes the accuracy and speed of bounding box regression. The experimental results show that, compared with the original model, the SOP-DETR model has increased the precision rate, recall rate and mAP@0.5 by 7.7, 3.3 and 4.5 percentage points respectively, while reducing the computational load by 30.4%, effectively enhancing the garbage detection performance in the complex underwater environment.

Key words: underwater garbage, object detection, RT-DETR, feature pyramid, loss function