Underwater object detection is crucial for marine engineering yet challenged by complex optical properties causing blurring, low contrast, and texture loss. Furthermore, deploying deep learning models on resource-constrained platforms necessitates balancing accuracy with latency. We propose a novel lightweight framework based on the Real-Time Detection Transformer (RT-DETR). First, we design the FasterFDBlock backbone, integrating Partial Convolution with Frequency-domain Dynamic Convolution. This utilizes frequency band modulation to adaptively suppress high-frequency noise and enhance edge details, optimizing feature extraction with reduced redundancy. Second, to mitigate small object detail loss, we introduce the AIFI-SEFN encoder, incorporating a Spatially Enhanced Feed-Forward Network to synthesize global semantic context with local spatial data. Third, a Multi-scale Feature Modulation (MFM) module is applied to dynamically weight deep semantic and shallow detailed features, bolstering robustness against scale variations and background interference. Experimental results on the UTDAC2020 dataset show our method achieves a mean Average Precision (mAP) of 72.1%, outperforming the baseline by 1.7%. Crucially, parameters and Floating Point Operations (GFLOPs) are reduced by 27.1% and 24.6%, respectively. With an inference speed of 72.6 FPS, this model offers a highly efficient solution for real-time underwater perception.
Cai et al. (Tue,) studied this question.