An Anchor-Free Detector With Channel-Based Prior and Bottom-Enhancement for Underwater Object Detection

The rapid and efficient detection of marine organisms through intelligent devices contributes to the marine economy. A vital stage is the accurate processing of the images obtained by vision sensors. However, more accurate object detectors usually attach large model sizes and have expensive computation costs. Meanwhile, for the selective absorption of light by water and the small size of underwater objects, the performance of existing detectors is not satisfactory. In this article, on the basis of anchor-free YOLOX-S, two lightweight modules are inserted to make the detection model more suitable for real-time object detection tasks in complex underwater environments. The attention-GB model does add almost no computational burden. It is employed to introduce the prior knowledge that the attenuation coefficients of red light, green light, and blue light in water are inconsistent. The bottom-enhancement module is applied to compensate the rich semantic information to the shallow layer to improve the small object detection accuracy. In addition, inspired by the idea of multiple instance learning (MIL), we put forward some strategies to reannotate the images with wrong labels in the dataset. The detection accuracy of YOLOX-U for holothurian is 2.2% AP higher than YOLOv8-S. Compared with other underwater object detectors, our proposed object detector achieves the best speed-accuracy tradeoff.

I. Introduction

Most of the traditional underwater operations are done by divers. Compared with divers, underwater robots, which are installed with multiple sensors [1], [2], [3], have a much longer and safer single operation time. Benefiting from the rapid development of object detection technology, underwater robots equipped with manipulators and vision sensors can replace frogmen to complete numerous close-range operations.