UCDN: A CenterNet-Based Dense Muti-scale Detection Fusion Net on Underwater Objects

Abstract

Underwater object detection is an inevitable task with urgent practical importance in the field of economic marine life. Due to the special underwater optical environment, most of the existing common detection algorithms are not capable of providing ideal results for underwater objects. For this reason, this paper proposes a constructive CenterNet-based underwater object detection model, named UCDN, accompanied by a detection strategy with pervasive applicability. Specifically, a detailed fusion module is created with the goal of filtering out interfering information. Meanwhile, we propose an exclusive idea of dense scale linking to fuse multi-scale features as much as possible. More importantly, we fuse our detection network with the Frankle-McCann Retinex algorithm to detect more objects obscured by the environment without increasing the training consumption. In addition to this, an efficient automatic sample balancing strategy is proposed, which is well suited to our detection situation. Finally, we evaluate our algorithm on underwater image datasets. The experiment results showed that the precision (mAP) of UCDN reached 87.46% which was higher than existing state-of-the-art land-based detection algorithms and underwater detection algorithms.