Detection of Water Surface Targets Based on Improved Deformable DETR

Abstract：ObjectiveThis study aims to propose an improved object detection algorithm based on Deformable DETR for intelligent recognition of water surface targets.Methods By substituting the original feature extraction network of Deformable DETR with a lightweight MobileNetV3 and introducing the CBAM attention mechanism module, the study achieves more efficient and robust detection of water surface targets. The effectiveness of the improved algorithm was verified through ablation experiments and horizontal comparative trials conducted on a self-constructed surface water target dataset and the publicly available ABOships dataset.ResultsAblation experiments conducted on both the self-constructed dataset and the ABOships dataset have demonstrated that the improved algorithm significantly reduces the model's parameter count and size to one-third of the original model. In terms of inference speed, there is an increase of 52.0% and 82.7% respectively on these datasets, while the mean Average Precision (mAP) at 0.5:0.95 has improved by 2.4% and 7.5%, and the training time has been reduced to merely 41.7% and 51.9% of the original algorithm, respectively. Further comparative tests of different algorithms conducted on the ABOships dataset underscore the superior performance of the proposed improved algorithm in both inference speed and detection accuracy. ConclusionsThis study has constructed a new water surface target dataset and made significant improvements to the Deformable DETR algorithm. These improvements have not only greatly increased the inference and training speed of the model but also enhanced its detection accuracy. This demonstrates the effectiveness and potential of DETR-class algorithms in the field of water surface target detection.