Toward a unified framework for feature enhancement-guided marine organism detection

Marine organism detection is a crucial technology for underwater autonomous robots, playing a pivotal role in enabling intelligent grasping and facilitating ocean exploration. However, the underwater images acquired by underwater robots through sensor devices have challenges such as low contrast, blur, and color cast. Additionally, the presence of various marine organism types and significant attitude variations further complicate the task of marine organism detection. We propose UEDNet, an innovative and integrated paradigm that combines visual enhancement and object detection tasks through an effective transformer-based feature enhancement module. Unlike conventional approaches that treat underwater image enhancement as a preliminary step, our framework adopts a multi-task joint learning strategy. This strategy allows for the effective sharing of enhanced features generated by the backbone module, promoting a comprehensive integration of weakened and enhanced features. This kind of integration plays a critical role in mitigating the detrimental impact that underperforming enhancement modules have on the detection module. Furthermore, we introduce an enhancement-supervised combination loss, which enables the detection module to handle varying degrees of underwater image degradation and reduces false detections and missed instances of marine organisms. UEDNet achieves a significantly high mean Average Precision (mAP) value of 79.81%, underscoring its robustness as a detection framework that bridges the gap between low-level underwater image enhancement and high-level marine object detection tasks.