U-DECN: End-to-End Underwater Object Detection ConvNet with Improved DeNoising Training

TL;DR

The paper tackles real-time underwater object detection under color cast noise by introducing U-DECN, a ConvNet encoder–decoder end-to-end detector. It combines a Hybrid Encoder, Two-Stage Bounding Box Refinement with Deformable Convolution in SIM, an Underwater Color DeNoising Query, and a Separate Contrastive DeNoising Forward to enhance convergence and robustness. On the DUO and RUOD benchmarks, U-DECN achieves 64.0 AP and 58.1 AP respectively, while delivering up to 21 FPS on NVIDIA AGX Orin (TensorRT FP16), outperforming state-of-the-art query-based detectors in deployment efficiency. The work includes comprehensive ablations that confirm the importance of each component, and provides code to facilitate adoption in underwater robotics applications.

Abstract

Underwater object detection has higher requirements of running speed and deployment efficiency for the detector due to its specific environmental challenges. NMS of two- or one-stage object detectors and transformer architecture of query-based end-to-end object detectors are not conducive to deployment on underwater embedded devices with limited processing power. As for the detrimental effect of underwater color cast noise, recent underwater object detectors make network architecture or training complex, which also hinders their application and deployment on unmanned underwater vehicles. In this paper, we propose the Underwater DECO with improved deNoising training (U-DECN), the query-based end-to-end object detector (with ConvNet encoder-decoder architecture) for underwater color cast noise that addresses the above problems. We integrate advanced technologies from DETR variants into DECO and design optimization methods specifically for the ConvNet architecture, including Deformable Convolution in SIM and Separate Contrastive DeNoising Forward methods. To address the underwater color cast noise issue, we propose an Underwater Color DeNoising Query method to improve the generalization of the model for the biased object feature information by different color cast noise. Our U-DECN, with ResNet-50 backbone, achieves the best 64.0 AP on DUO and the best 58.1 AP on RUOD, and 21 FPS (5 times faster than Deformable DETR and DINO 4 FPS) on NVIDIA AGX Orin by TensorRT FP16, outperforming the other state-of-the-art query-based end-to-end object detectors. The code is available at https://github.com/LEFTeyex/U-DECN.

Figures (6)

• Figure 1: U-DECN architecture for color cast noise.
• Figure 2: The framework of our proposed U-DECN which is based on the ConvNet encoder-decoder architecture. The hybrid encoder fuses three scale features in F4 by CNN-based Cross-scale Feature Fusion (CCFF)rtdetr. The top-K features are selected to initialize positional queries dynamically, while content queries are kept as learnable parameters. The color denoising query is encoded from the input image. U-DECN also contains a Contrastive DeNoising (CDN) part with positive and negative samples, which is trained by Separate CDN Forward.
• Figure 3: The pipeline of deformable convolution for the query feature misalignment issue in the self-interaction module (SIM).
• Figure 4: Training convergence curves of DECO deco, DETR detr, and U-DECN with ResNet-50 resnet.
• Figure 5: Comparisons of detection results. The yellow circles contain duplicates, misses, and false detection issues.