Underwater Object Detection Based on Transformer Query Optimization

Due to the complex and changeable underwater environment, the quality of underwater images is significantly degraded by uneven lighting. When generic object detection algorithms are directly applied to underwater scenes, the detection results are not ideal. To address the above issues, we propose a novel underwater detector with multi-scale feature enhancement and query optimization. Specifically, we propose a Conv-Trans Layer, which enables the model to effectively capture different levels of information in the image through CNN, and captures the dependencies between different positions in the sequence data through Transformer, enabling the model to process Sequence information over longer distances. The effective combination of the two can enhance the representation of multi-scale features. Then, we develop a Positional Queries Enhancement Module to optimize the spatial prior of positional queries, thereby speeding up the convergence of the network. In addition, we also construct a new underwater object detection dataset called UODD2023, which includes more than 20,000 underwater images for training and validation. Extensive experiments on COCO and UODD2023 show that our method outperforms other methods in terms of robustness and accuracy.