A review of underwater object detection methods based on deep learning

Abstract

Underwater object detection, a fundamental yet challenging task in ocean engineering, has been greatly advanced by the rapid development of deep learning. This paper provides a comprehensive review of recent deep learning–based methods for underwater object detection. Existing approaches are systematically categorized into seven groups: transfer of general object detection methods, feature fusion, image enhancement, feature enhancement, domain generalization, transformer-based models, and other emerging techniques. Representative algorithms within each category are analyzed and compared. Furthermore, the paper introduces commonly used evaluation metrics for both object detection and image quality assessment, and discusses four major challenges in the field: image feature degradation, variability in object scales, generalization capability, and the adaptability of models to robotic platforms. In addition, publicly available underwater datasets are reviewed and the performance of various methods on these datasets is summarized. Finally, based on the above synthesis, potential research directions and future perspectives for underwater object detection are highlighted. Compared to previous review articles, this paper proposes a classification scheme based on the unique characteristics of the underwater environment. Starting from the specific challenges being addressed, it systematically summarizes and categorizes existing methods, providing valuable insights for future methodological improvements. The article not only describes various techniques but also conducts a systematic comparison of representative literature through multiple structured tables. It reviews emerging technologies such as Transformer, Mamba, multi-modal fusion, and weakly-supervised learning, reflecting the latest trends in the field. Particularly, it introduces the application of Mamba in underwater image enhancement and object detection. Furthermore, it systematically summarizes commonly used evaluation metrics for object detection as well as underwater image quality assessment metrics.