UOD-FPN: An Effective Feature Pyramid Network for Underwater Object Detection

Abstract

Optical images captured in underwater environments often suffer from degradation problems such as color deviation, low contrast, and haze effects. Faced with this situation, traditional feature pyramid networks (FPNs) have difficulty effectively integrating the multi-scale features of underwater objects, which limits the performance of object detectors in underwater environments. Therefore, this paper proposes a novel underwater object detection feature pyramid network (UOD-FPN), which can effectively enhance the ability to represent underwater object features. This network consists of three different modules. First, the inter-layer feature fusion module (IFM) is capable of generating multiple aggregated feature layers with different receptive fields. Second, the feature adaptive weighted re-layering module (FAWRM) introduces attention mechanisms to guide finer re-layering of the aggregated feature layer. Finally, the feature-selective fusion module (FFM) achieves more effective feature fusion by using cross-correlation responses between different feature layers as guidance, enhancing the ability to express multi-level integrated features. Additionally, the UOD-FPN proposed in this paper can be applied to most FPN-based object detection models. The experiments demonstrate that UOD-FPN can significantly enhance the performance of one-stage and two-stage object detectors on the underwater datasets DUO and RUOD.