A Lightweight Object Detector With Deformable Upsampling for Marine Organism Detection

Marine organism detection is a significant topic in the rational development and utilization of ocean resources. Due to the low computational ability of underwater vehicles, large-scale object detection models cannot be applied to them. In this paper, firstly, a lightweight feature extraction network named Mobile-bone is adopted, which not only significantly reduces parameters but also combines the advantages of convolutional neural networks (CNNS) and vision transformers (ViTs) to learn global representations. Secondly, we put forward a novel upsampling method named deformable upsampling for feature fusion networks. Our proposed deformable upsampling is a generalization-effective upsampling operation that leverages semantic alignment rather than spatial alignment to reduce the error in the upsampling process. Experimental results indicate that deformable upsampling is appropriate for diverse feature fusion networks and significantly boosts the precision of underwater object detectors by only increasing 0.39 M parameters. Finally, our proposed detector has promising detection accuracy on the underwater open dataset, and it has also performed exceptionally well when ported to the embedded device for detecting marine organisms in real-world scenarios. Code and models about DU-MobileYOLO are available at: https://github.com/ZERO-SPACE-X/ DU-MobileYOLO.